Apple Patent | Bit Stream Structure For Compressed Point Cloud Data

Patent: Bit Stream Structure For Compressed Point Cloud Data

Publication Number: 20200021847

Publication Date: 20200116

Applicants: Apple

Abstract

A system comprises an encoder configured to compress attribute information and/or spatial information for a point cloud and/or a decoder configured to decompress compressed attribute and/or spatial information for the point cloud. To compress the attribute and/or spatial information, the encoder is configured to convert a point cloud into an image based representation. Also, the decoder is configured to generate a decompressed point cloud based on an image based representation of a point cloud. In some embodiments, a bit stream structure may be used to communicate compressed point cloud data. The bit stream structure may include point cloud compression network abstraction layer (PCCNAL) units that enable use of groups of frames (GOFs), frame, and sub-frame signaling of patch information. Such a bit stream structure may permit low delay streaming and random access reconstruction of point clouds amongst other applications.

PRIORITY CLAIM

[0001] This application claims benefit of priority to U.S. Provisional Application Ser. No. 62/697,369, entitled “Bit Stream Structure for Compressed Point Cloud Data”, filed Jul. 12, 2018, and which is incorporated herein by reference in its entirety.

BACKGROUND

Technical Field

[0002] This disclosure relates generally to compression and decompression of point clouds comprising a plurality of points, each having associated spatial information and attribute information.

Description of the Related Art

[0003] Various types of sensors, such as light detection and ranging (LIDAR) systems, 3-D-cameras, 3-D scanners, etc. may capture data indicating positions of points in three dimensional space, for example positions in the X, Y, and Z planes. Also, such systems may further capture attribute information in addition to spatial information for the respective points, such as color information (e.g. RGB values), texture information, intensity attributes, reflectivity attributes, motion related attributes, modality attributes, or various other attributes. In some circumstances, additional attributes may be assigned to the respective points, such as a time-stamp when the point was captured. Points captured by such sensors may make up a “point cloud” comprising a set of points each having associated spatial information and one or more associated attributes. In some circumstances, a point cloud may include thousands of points, hundreds of thousands of points, millions of points, or even more points. Also, in some circumstances, point clouds may be generated, for example in software, as opposed to being captured by one or more sensors. In either case, such point clouds may include large amounts of data and may be costly and time-consuming to store and transmit.

SUMMARY OF EMBODIMENTS

[0004] In some embodiments, a system includes one or more sensors configured to capture points that collectively make up a point cloud, wherein each of the points comprises spatial information identifying a spatial location of the respective point and attribute information defining one or more attributes associated with the respective point.

[0005] The system also includes an encoder configured to compress the attribute and/or spatial information of the points. To compress the attribute and/or spatial information, the encoder is configured to determine, for the point cloud, a plurality of patches, each corresponding to portions of the point cloud. The encoder is further configured to, for each patch, generate a patch image comprising the set of points corresponding to the patch projected onto a patch plane and generate another patch image comprising geometry information, such as depth information, for the set of points corresponding to the patch, wherein the geometry information comprises depths of the points in a direction perpendicular to the patch plane.

[0006] For example, the geometry patch image corresponding to the patch projected onto a patch plane may depict the points of the point cloud included in the patch in two directions, such as an X and Y direction. The points of the point cloud may be projected onto a patch plane approximately perpendicular to a normal vector, normal to a surface of the point cloud at the location of the patch. Also, for example, the geometry patch image comprising depth information for the set of points included in the patch may depict depth information, such as depth distances in a Z direction. To depict the depth information, the geometry patch image may include a parameter that varies in intensity based on the depth of points in the point cloud at a particular location in the patch image. For example, the geometry patch image depicting depth information may have a same shape as the attribute patch image representing attributes of points projected onto the patch plane. However, the geometry information patch image may be an image comprising image attributes, such as one or more colors, that vary in intensity based on depth, wherein the intensity of the one or more image attributes corresponds to a depth of a corresponding point of the point cloud at a location in the geometry patch image where the image attribute is displayed in the geometry patch image depicting depth. For example, points that are closer to the patch plane may be encoded as darker values in the patch image depicting depth and points that are further away from the patch plane may be encoded as lighter values in the patch image depicting depth, for example in a monochromatic patch image depicting depth. Thus, the depth information patch image when aligned with other patch images representing attribute values for points projected onto the patch plane may indicate the relative depths of the points projected onto the patch plane, based on respective image attribute intensities at locations in the geometry patch image that correspond to locations of the points in the other patch images comprising point cloud points projected onto the patch plane.

[0007] The encoder is further configured to pack generated patch images (including a geometry patch image and one or more additional patch images for one or more other attributes such as colors, textures, reflectances, etc.) for each of the determined patches into one or more image frames. Also, the encoder is configured to provide the one or more packed image frames to a video encoding component (which may be included in the encoder or may be a separate video encoding component). Additionally, the encoder is configured to insert one or more point cloud compression network abstraction layer units (PCCNAL units) in a bit stream comprising the video encoded packed image frames.

[0008] In some embodiments, the PCCNAL units may be inserted anywhere in the bit stream and may include a sequence of leading bits, trailing bits, or both that make the PCCNAL units easily identifiable by a decoder. In some embodiments, the PCCNAL units may further include organizational information that enables a decoder to quickly identify and decode a given set of patch images packed into one or more of the video encoded image frames of the bit stream. For example, the PCCNAL units may enable a decoder to identify related patch images corresponding to a particular portion of a point cloud to enable low delay reconstruction of the particular portion of the point cloud, wherein necessary patch images needed to reconstruct the portion of the point cloud are identified in the bit stream using the organization information included in the PCCNAL units.

[0009] In some embodiments, the encoder may utilize a video encoding component in accordance with the High Efficiency Video Coding (HEVC/H.265) standard or other suitable standards such as, the Advanced Video Coding (AVC/H.264) standard, the AOMedia VIDEO 1 (AV1) video coding format produced by the Alliance for Open Media (AOM), etc. In some embodiments, the encoder may utilize an image encoder in accordance with a Motion Picture Experts Group (MPEG), a Joint Photography Experts Group (JPEG) standard, an International Telecommunication Union-Telecommunication standard (e.g. ITU-T standard), etc.

[0010] In some embodiments, a decoder is configured to receive one or more encoded image frames comprising patch images for a plurality of patches of a compressed point cloud, wherein, for each patch, the one or more encoded image frames comprise: a patch image comprising a set of points of the patch projected onto a patch plane and a patch image comprising depth information for the set of points of the patch, wherein the depth information indicates depths of the points of the patch in a direction perpendicular to the patch plane. In some embodiments, a depth patch image may be packed into an image frame with other attribute patch images. For example, a decoder may receive one or more image frames comprising packed patch images as generated by the encoder described above. A bit stream comprising the video encoded image frames may further included one or more point cloud compression network abstraction layer units (PCCNAL units). In some embodiments, a decoder may scan a bit stream for PCCNAL units and may further utilize organization information included in the PCCNAL units in order to determine which video encoded image frames or portions of the video encoded image frames to decoder, and in what order to decode them.

[0011] Furthermore, the decoder is further configured to video decode the one or more identified video encoded image frames comprising the patch images according to the determined order. In some embodiments, the decoder may utilize a video decoder in accordance with the High Efficiency Video Coding (HEVC) standard or other suitable standards such as, the Advanced Video Coding (AVC) standard, the AOMedia Video 1 (AV1) video coding format, etc. In some embodiments, the decoder may utilize an image decoder in accordance with a Motion Picture Experts Group (MPEG) or a Joint Photography Experts Group (JPEG) standard, etc.

[0012] The decoder is further configured to determine, for each patch, spatial information for the set of points of the patch based, at least in part, on the attribute patch image comprising the set of points of the patch projected onto the patch plane and the geometry patch image comprising the depth information for the set of points of the patch, and generate a reconstructed version of the compressed point cloud based, at least in part, on the determined spatial information for the plurality of patches and the attribute information included in the patches.

[0013] In some embodiments, a method includes receiving data for a point cloud comprising a plurality of points that make up the point cloud, wherein respective ones of the points comprise spatial information for the point and attribute information for the point. The method further includes compressing the point cloud data, wherein compressing the point cloud data comprises determining, for the point cloud, a plurality of patches each corresponding to portions of the point cloud and packing generated patch images for the determined patches into one or more image frames. Also, compression the point cloud data comprises encoding the one or more image frames into a bit stream comprising point cloud compression network abstraction layer (PCCNAL) units, wherein the PCCNAL units indicate locations of one or more of the patch images in the bit stream.

[0014] In some embodiments, a method includes receiving an encoded bit stream comprising video encoded image frames comprising patch images packed into respective ones of the video encoded image frames and comprising one or more point cloud compression network abstraction layer (PCCNAL) units in the bit stream. The method also includes identifying one or more of the PCCNAL units in the bit stream and determining an order in which to video decode at least some of the patch images packed into the video encoded image frames based on information indicated in the one or more PCCNAL units. Additionally, the method includes video decoding a set of video encoded image frames comprising the at least some patch images according to the order determined based on the information indicated in the one or more PCCNAL units.

[0015] In some embodiments, a non-transitory computer-readable medium stores program instructions that, when executed by one or more processors, cause the one or more processors to implement an encoder as described herein to compress geometry and attribute information of a point cloud.

[0016] In some embodiments, a non-transitory computer-readable medium stores program instructions that, when executed by one or more processors, cause the one or more processors to implement a decoder as described herein to decompress geometry and attribute information of a point cloud.

BRIEF DESCRIPTION OF THE DRAWINGS

[0017] FIG. 1 illustrates a system comprising a sensor that captures information for points of a point cloud and an encoder that compresses spatial information and attribute information of the point cloud, where the compressed spatial and attribute information is sent to a decoder, according to some embodiments.

[0018] FIG. 2A illustrates components of an encoder for encoding intra point cloud frames, according to some embodiments.

[0019] FIG. 2B illustrates components of a decoder for decoding intra point cloud frames, according to some embodiments.

[0020] FIG. 2C illustrates components of an encoder for encoding inter point cloud frames, according to some embodiments.

[0021] FIG. 2D illustrates components of a decoder for decoding inter point cloud frames, according to some embodiments.

[0022] FIG. 3A illustrates an example patch segmentation process, according to some embodiments.

[0023] FIG. 3B illustrates an example image frame comprising packed patch images and padded portions, according to some embodiments.

[0024] FIG. 3C illustrates an example image frame comprising patch portions and padded portions, according to some embodiments.

[0025] FIG. 3D illustrates a point cloud being projected onto multiple projections, according to some embodiments.

[0026] FIG. 3E illustrates a point cloud being projected onto multiple parallel projections, according to some embodiments.

[0027] FIG. 4A illustrates components of an encoder for encoding intra point cloud frames with color conversion, according to some embodiments.

[0028] FIG. 4B illustrates components of an encoder for encoding inter point cloud frames with color conversion, according to some embodiments.

[0029] FIG. 4C illustrates components of a closed-loop color conversion module, according to some embodiments.

[0030] FIG. 4D illustrates an example process for determining a quality metric for a point cloud upon which an operation has been performed, according to some embodiments.

[0031] FIG. 5A illustrates components of an encoder that includes geometry, texture, and/or other attribute downscaling, according to some embodiments.

[0032] FIG. 5B illustrates components of a decoder that includes geometry, texture, and/or other attribute upscaling, according to some embodiments.

[0033] FIG. 5C illustrates rescaling from the perspective of an encoder, according to some embodiments.

[0034] FIG. 5D illustrates rescaling from the perspective of a decoder, according to some embodiments.

[0035] FIG. 5E illustrates an example open loop rescaling, according to some embodiments.

[0036] FIG. 5F illustrates an example closed loop rescaling, according to some embodiments.

[0037] FIG. 5G illustrates an example closed loop rescaling with multiple attribute layers, according to some embodiments.

[0038] FIG. 5H illustrates an example of video level spatiotemporal scaling, according to some embodiments.

[0039] FIG. 5I illustrates an example closed loop rescaling with spatiotemporal scaling, according to some embodiments.

[0040] FIG. 5J illustrates a process of encoding/compressing image frames of a point cloud using down-scaling, according to some embodiments.

[0041] FIG. 5K illustrates a process of determining to down-scaling image frames using open-loop or closed-loop down-scaling, according to some embodiments.

[0042] FIG. 5L illustrates a process of decoding/decompressing image frames of a point cloud using up-scaling, according to some embodiments.

[0043] FIG. 6A illustrates components of an encoder that further includes pre-video compression texture processing and/or filtering and pre video compression geometry processing/filtering, according to some embodiments.

[0044] FIG. 6B illustrates components of a decoder that further includes post video decompression texture processing and/or filtering and post video decompression geometry processing/filtering, according to some embodiments.

[0045] FIG. 6C illustrates, a bit stream structure for a compressed point cloud, according to some embodiments.

[0046] FIG. 6D illustrates a process for generating video encoded image frames for patches of a point cloud taking into account relationship information between the patches packed into the image frames, according to some embodiments.

[0047] FIG. 6E illustrates a process for generating video encoded image frames taking into account pooled distortion for a set of patches corresponding to a same set of points, according to some embodiments.

[0048] FIG. 6F illustrates a process for generating video encoded image frames taking into account patch edges, according to some embodiments.

[0049] FIG. 6G illustrates a process for reconstructing a point cloud based on video encoded image frames comprising patches of the point cloud, wherein relationship information between the patches packed into the image frames is taken into account, according to some embodiments.

[0050] FIG. 6H illustrates a process of upscaling a patch image included in an image frame taking into account edges of the patch image determined based on received or determined relationship information for the patches, according to some embodiments.

[0051] FIG. 6I illustrates an example application where an attribute plane is up-scaled using its corresponding geometry information and the geometry extracted edges, according to some embodiments.

[0052] FIG. 7A illustrates an example of a point cloud compression network abstraction layer (PCCNAL) unit based bit stream, according to some embodiments.

[0053] FIG. 7B illustrates an example of a PCCNAL units grouped by picture order count (POC), according to some embodiments.

[0054] FIG. 7C illustrates an example of a PCCNAL unit grouped by type, according to some embodiments.

[0055] FIG. 7D illustrates a process of encoding a bit stream that includes PCCNAL units, according to some embodiments.

[0056] FIG. 7E illustrates a process of decoding a bit stream that includes PCCNAL units, according to some embodiments.

[0057] FIG. 8A illustrates a process for compressing attribute and spatial information of a point cloud, according to some embodiments.

[0058] FIG. 8B illustrates a process for decompressing attribute and spatial information of a point cloud, according to some embodiments.

[0059] FIG. 8C illustrates patch images being generated and packed into an image frame to compress attribute and spatial information of a point cloud, according to some embodiments.

[0060] FIG. 9 illustrates patch images being generated and packed into an image frame to compress attribute and spatial information of a moving or changing point cloud, according to some embodiments.

[0061] FIG. 10 illustrates a decoder receiving image frames comprising patch images, patch information, and an occupancy map, and generating a decompressed representation of a point cloud, according to some embodiments.

[0062] FIG. 11A illustrates an encoder, adjusting encoding based on one or more masks for a point cloud, according to some embodiments.

[0063] FIG. 11B illustrates a decoder, adjusting decoding based on one or more masks for a point cloud, according to some embodiments.

[0064] FIG. 12A illustrates more detail regarding compression of an occupancy map, according to some embodiments.

[0065] FIG. 12B illustrates example blocks and traversal patterns for compressing an occupancy map, according to some embodiments.

[0066] FIG. 12C illustrates more detail regarding compression of an occupancy map, according to some embodiments.

[0067] FIG. 13 illustrates compressed point cloud information being used in a 3-D telepresence application, according to some embodiments.

[0068] FIG. 14 illustrates compressed point cloud information being used in a virtual reality application, according to some embodiments.

[0069] FIG. 15 illustrates an example computer system that may implement an encoder or decoder, according to some embodiments.

[0070] This specification includes references to “one embodiment” or “an embodiment.” The appearances of the phrases “in one embodiment” or “in an embodiment” do not necessarily refer to the same embodiment. Particular features, structures, or characteristics may be combined in any suitable manner consistent with this disclosure.

[0071] “Comprising.” This term is open-ended. As used in the appended claims, this term does not foreclose additional structure or steps. Consider a claim that recites: “An apparatus comprising one or more processor units … .” Such a claim does not foreclose the apparatus from including additional components (e.g., a network interface unit, graphics circuitry, etc.).

[0072] “Configured To.” Various units, circuits, or other components may be described or claimed as “configured to” perform a task or tasks. In such contexts, “configured to” is used to connote structure by indicating that the units/circuits/components include structure (e.g., circuitry) that performs those task or tasks during operation. As such, the unit/circuit/component can be said to be configured to perform the task even when the specified unit/circuit/component is not currently operational (e.g., is not on). The units/circuits/components used with the “configured to” language include hardware–for example, circuits, memory storing program instructions executable to implement the operation, etc. Reciting that a unit/circuit/component is “configured to” perform one or more tasks is expressly intended not to invoke 35 U.S.C. .sctn. 112(f), for that unit/circuit/component. Additionally, “configured to” can include generic structure (e.g., generic circuitry) that is manipulated by software and/or firmware (e.g., an FPGA or a general-purpose processor executing software) to operate in manner that is capable of performing the task(s) at issue. “Configure to” may also include adapting a manufacturing process (e.g., a semiconductor fabrication facility) to fabricate devices (e.g., integrated circuits) that are adapted to implement or perform one or more tasks.

[0073] “First,” “Second,” etc. As used herein, these terms are used as labels for nouns that they precede, and do not imply any type of ordering (e.g., spatial, temporal, logical, etc.). For example, a buffer circuit may be described herein as performing write operations for “first” and “second” values. The terms “first” and “second” do not necessarily imply that the first value must be written before the second value.

[0074] “Based On.” As used herein, this term is used to describe one or more factors that affect a determination. This term does not foreclose additional factors that may affect a determination. That is, a determination may be solely based on those factors or based, at least in part, on those factors. Consider the phrase “determine A based on B.” While in this case, B is a factor that affects the determination of A, such a phrase does not foreclose the determination of A from also being based on C. In other instances, A may be determined based solely on B.

DETAILED DESCRIPTION

[0075] As data acquisition and display technologies have become more advanced, the ability to capture point clouds comprising thousands or millions of points in 2-D or 3-D space, such as via LIDAR systems, has increased. Also, the development of advanced display technologies, such as virtual reality or augmented reality systems, has increased potential uses for point clouds. However, point cloud files are often very large and may be costly and time-consuming to store and transmit. For example, communication of point clouds over private or public networks, such as the Internet, may require considerable amounts of time and/or network resources, such that some uses of point cloud data, such as real-time uses, may be limited. Also, storage requirements of point cloud files may consume a significant amount of storage capacity of devices storing the point cloud files, which may also limit potential applications for using point cloud data.

[0076] In some embodiments, an encoder may be used to generate a compressed point cloud to reduce costs and time associated with storing and transmitting large point cloud files. In some embodiments, a system may include an encoder that compresses attribute and/or spatial information of a point cloud file such that the point cloud file may be stored and transmitted more quickly than non-compressed point clouds and in a manner that the point cloud file may occupy less storage space than non-compressed point clouds. In some embodiments, compression of attributes of points in a point cloud may enable a point cloud to be communicated over a network in real-time or in near real-time. For example, a system may include a sensor that captures attribute information about points in an environment where the sensor is located, wherein the captured points and corresponding attributes make up a point cloud. The system may also include an encoder that compresses the captured point cloud attribute information. The compressed attribute information of the point cloud may be sent over a network in real-time or near real-time to a decoder that decompresses the compressed attribute information of the point cloud. The decompressed point cloud may be further processed, for example to make a control decision based on the surrounding environment at the location of the sensor. The control decision may then be communicated back to a device at or near the location of the sensor, wherein the device receiving the control decision implements the control decision in real-time or near real-time. In some embodiments, the decoder may be associated with an augmented reality system and the decompressed attribute information may be displayed or otherwise used by the augmented reality system. In some embodiments, compressed attribute information for a point cloud may be sent with compressed spatial information for points of the point cloud. In other embodiments, spatial information and attribute information may be separately encoded and/or separately transmitted to a decoder.

[0077] In some embodiments, a system may include a decoder that receives one or more sets of point cloud data comprising compressed attribute information via a network from a remote server or other storage device that stores the one or more point cloud files. For example, a 3-D display, a holographic display, or a head-mounted display may be manipulated in real-time or near real-time to show different portions of a virtual world represented by point clouds. In order to update the 3-D display, the holographic display, or the head-mounted display, a system associated with the decoder may request point cloud data from the remote server based on user manipulations of the displays, and the point cloud data may be transmitted from the remote server to the decoder and decoded by the decoder in real-time or near real-time. The displays may then be updated with updated point cloud data responsive to the user manipulations, such as updated point attributes.

[0078] In some embodiments, a system, may include one or more LIDAR systems, 3-D cameras, 3-D scanners, etc., and such sensor devices may capture spatial information, such as X, Y, and Z coordinates for points in a view of the sensor devices. In some embodiments, the spatial information may be relative to a local coordinate system or may be relative to a global coordinate system (for example, a Cartesian coordinate system may have a fixed reference point, such as a fixed point on the earth, or may have a non-fixed local reference point, such as a sensor location).

[0079] In some embodiments, such sensors may also capture attribute information for one or more points, such as color attributes, texture attributes, reflectivity attributes, velocity attributes, acceleration attributes, time attributes, modalities, and/or various other attributes. In some embodiments, other sensors, in addition to LIDAR systems, 3-D cameras, 3-D scanners, etc., may capture attribute information to be included in a point cloud. For example, in some embodiments, a gyroscope or accelerometer, may capture motion information to be included in a point cloud as an attribute associated with one or more points of the point cloud. For example, a vehicle equipped with a LIDAR system, a 3-D camera, or a 3-D scanner may include the vehicle’s direction and speed in a point cloud captured by the LIDAR system, the 3-D camera, or the 3-D scanner. For example, when points in a view of the vehicle are captured they may be included in a point cloud, wherein the point cloud includes the captured points and associated motion information corresponding to a state of the vehicle when the points were captured.

Example System Arrangement

[0080] FIG. 1 illustrates a system comprising a sensor that captures information for points of a point cloud and an encoder that compresses attribute information of the point cloud, where the compressed attribute information is sent to a decoder, according to some embodiments.

[0081] System 100 includes sensor 102 and encoder 104. Sensor 102 captures a point cloud 110 comprising points representing structure 106 in view 108 of sensor 102. For example, in some embodiments, structure 106 may be a mountain range, a building, a sign, an environment surrounding a street, or any other type of structure. In some embodiments, a captured point cloud, such as captured point cloud 110, may include spatial and attribute information for the points included in the point cloud. For example, point A of captured point cloud 110 comprises X, Y, Z coordinates and attributes 1, 2, and 3. In some embodiments, attributes of a point may include attributes such as R, G, B color values, a velocity at the point, an acceleration at the point, a reflectance of the structure at the point, a time stamp indicating when the point was captured, a string-value indicating a modality when the point was captured, for example “walking”, or other attributes. The captured point cloud 110 may be provided to encoder 104, wherein encoder 104 generates a compressed version of the point cloud (compressed attribute information 112) that is transmitted via network 114 to decoder 116. In some embodiments, a compressed version of the point cloud, such as compressed attribute information 112, may be included in a common compressed point cloud that also includes compressed spatial information for the points of the point cloud or, in some embodiments, compressed spatial information and compressed attribute information may be communicated as separate sets of data.

[0082] In some embodiments, encoder 104 may be integrated with sensor 102. For example, encoder 104 may be implemented in hardware or software included in a sensor device, such as sensor 102. In other embodiments, encoder 104 may be implemented on a separate computing device that is proximate to sensor 102.

Example Intra-Frame Encoder

[0083] FIG. 2A illustrates components of an encoder for encoding intra point cloud frames, according to some embodiments. In some embodiments, the encoder described above in regard to FIG. 1 may operate in a similar manner as encoder 200 described in FIG. 2A and encoder 250 described in FIG. 2C.

[0084] The encoder 200 receives uncompressed point cloud 202 and generates compressed point cloud information 204. In some embodiments, an encoder, such as encoder 200, may receive the uncompressed point cloud 202 from a sensor, such as sensor 102 illustrated in FIG. 1, or, in some embodiments, may receive the uncompressed point cloud 202 from another source, such as a graphics generation component that generates the uncompressed point cloud in software, as an example.

[0085] In some embodiments, an encoder, such as encoder 200, includes decomposition into patches module 206, packing module 208, spatial image generation module 210, texture image generation module 212, and attribute information generation module 214. In some embodiments, an encoder, such as encoder 200, also includes image frame padding module 216, video compression module 218 and multiplexer 224. In addition, in some embodiments an encoder, such as encoder 200, may include an occupancy map compression module, such as occupancy map compression module 220, and an auxiliary patch information compression module, such as auxiliary patch information compression module 222. In some embodiments, an encoder, such as encoder 200, converts a 3D point cloud into an image-based representation along with some meta data (e.g., occupancy map and patch info) necessary to convert the compressed point cloud back into a decompressed point cloud.

[0086] In some embodiments, the conversion process decomposes the point cloud into a set of patches (e.g., a patch is defined as a contiguous subset of the surface described by the point cloud), which may be overlapping or not, such that each patch may be described by a depth field with respect to a plane in 2D space. More details about the patch decomposition process are provided above with regard to FIGS. 3A-3C.

[0087] After or in conjunction with the patches being determined for the point cloud being compressed, a 2D sampling process is performed in planes associated with the patches. The 2D sampling process may be applied in order to approximate each patch with a uniformly sampled point cloud, which may be stored as a set of 2D patch images describing the geometry/texture/attributes of the point cloud at the patch location. The “Packing” module 208 may store the 2D patch images associated with the patches in a single (or multiple) 2D images, referred to herein as “image frames” or “video image frames.” In some embodiments, a packing module, such as packing module 208, may pack the 2D patch images such that the packed 2D patch images do not overlap (even though an outer bounding box for one patch image may overlap an outer bounding box for another patch image). Also, the packing module may pack the 2D patch images in a way that minimizes non-used images pixels of the image frame.

[0088] In some embodiments, “Geometry/Texture/Attribute generation” modules, such as modules 210, 212, and 214, generate 2D patch images associated with the geometry/texture/attributes, respectively, of the point cloud at a given patch location. As noted before, a packing process, such as performed by packing module 208, may leave some empty spaces between 2D patch images packed in an image frame. Also, a padding module, such as image frame padding module 216, may fill in such areas in order to generate an image frame that may be suited for 2D video and image codecs.

[0089] In some embodiments, an occupancy map (e.g., binary information describing for each pixel or block of pixels whether the pixel or block of pixels are padded or not) may be generated and compressed, for example by occupancy map compression module 220. The occupancy map may be sent to a decoder to enable the decoder to distinguish between padded and non-padded pixels of an image frame.

[0090] Note that other metadata associated with patches may also be sent to a decoder for use in the decompression process. For example, patch information indicating sizes and shapes of patches determined for the point cloud and packed in an image frame may be generated and/or encoded by an auxiliary patch-information compression module, such as auxiliary patch-information compression module 222. In some embodiments one or more image frames may be encoded by a video encoder, such as video compression module 218. In some embodiments, a video encoder, such as video compression module 218, may operate in accordance with the High Efficiency Video Coding (HEVC) standard or other suitable video encoding standard. In some embodiments, encoded video images, encoded occupancy map information, and encoded auxiliary patch information may be multiplexed by a multiplexer, such as multiplexer 224, and provided to a recipient as compressed point cloud information, such as compressed point cloud information 204.

[0091] In some embodiments, an occupancy map may be encoded and decoded by a video compression module, such as video compression module 218. This may be done at an encoder, such as encoder 200, such that the encoder has an accurate representation of what the occupancy map will look like when decoded by a decoder. Also, variations in image frames due to lossy compression and decompression may be accounted for by an occupancy map compression module, such as occupancy map compression module 220, when determining an occupancy map for an image frame. In some embodiments, various techniques may be used to further compress an occupancy map, such as described in FIGS. 12A-12B.

Example Intra-Frame Decoder

[0092] FIG. 2B illustrates components of a decoder for decoding intra point cloud frames, according to some embodiments. Decoder 230 receives compressed point cloud information 204, which may be the same compressed point cloud information 204 generated by encoder 200. Decoder 230 generates reconstructed point cloud 246 based on receiving the compressed point cloud information 204.

[0093] In some embodiments, a decoder, such as decoder 230, includes a de-multiplexer 232, a video decompression module 234, an occupancy map decompression module 236, and an auxiliary patch-information decompression module 238. Additionally a decoder, such as decoder 230 includes a point cloud generation module 240, which reconstructs a point cloud based on patch images included in one or more image frames included in the received compressed point cloud information, such as compressed point cloud information 204. In some embodiments, a decoder, such as decoder 203, further comprises a smoothing filter, such as smoothing filter 244. In some embodiments, a smoothing filter may smooth incongruences at edges of patches, wherein data included in patch images for the patches has been used by the point cloud generation module to recreate a point cloud from the patch images for the patches. In some embodiments, a smoothing filter may be applied to the pixels located on the patch boundaries to alleviate the distortions that may be caused by the compression/decompression process.

Example Inter-Frame Encoder

[0094] FIG. 2C illustrates components of an encoder for encoding inter point cloud frames, according to some embodiments. An inter point cloud encoder, such as inter point cloud encoder 250, may encode an image frame, while considering one or more previously encoded/decoded image frames as references.

[0095] In some embodiments, an encoder for inter point cloud frames, such as encoder 250, includes a point cloud re-sampling module 252, a 3-D motion compensation and delta vector prediction module 254, a spatial image generation module 256, a texture image generation module 258, and an attribute image generation module 260. In some embodiments, an encoder for inter point cloud frames, such as encoder 250, may also include an image padding module 262 and a video compression module 264. An encoder for inter point cloud frames, such as encoder 250, may generate compressed point cloud information, such as compressed point cloud information 266. In some embodiments, the compressed point cloud information may reference point cloud information previously encoded by the encoder, such as information from or derived from one or more reference image frames. In this way an encoder for inter point cloud frames, such as encoder 250, may generate more compact compressed point cloud information by not repeating information included in a reference image frame, and instead communicating differences between the reference frames and a current state of the point cloud.

[0096] In some embodiments, an encoder, such as encoder 250, may be combined with or share modules with an intra point cloud frame encoder, such as encoder 200. In some embodiments, a point cloud re-sampling module, such as point cloud re-sampling module 252, may resample points in an input point cloud image frame in order to determine a one-to-one mapping between points in patches of the current image frame and points in patches of a reference image frame for the point cloud. In some embodiments, a 3D motion compensation & delta vector prediction module, such as a 3D motion compensation & delta vector prediction module 254, may apply a temporal prediction to the geometry/texture/attributes of the resampled points of the patches. The prediction residuals may be stored into images, which may be padded and compressed by using video/image codecs. In regard to spatial changes for points of the patches between the reference frame and a current frame, a 3D motion compensation & delta vector prediction module 254, may determine respective vectors for each of the points indicating how the points moved from the reference frame to the current frame. A 3D motion compensation & delta vector prediction module 254, may then encode the motion vectors using different image parameters. For example, changes in the X direction for a point may be represented by an amount of red included at the point in a patch image that includes the point. In a similar manner, changes in the Y direction for a point may be represented by an amount of blue included at the point in a patch image that includes the point. Also, in a similar manner, changes in the Z direction for a point may be represented by an amount of green included at the point in a patch image that includes the point. In some embodiments, other characteristics of an image included in a patch image may be adjusted to indicate motion of points included in the patch between a reference frame for the patch and a current frame for the patch.

Example Inter-Frame Decoder

[0097] FIG. 2D illustrates components of a decoder for decoding inter point cloud frames, according to some embodiments. In some embodiments, a decoder, such as decoder 280, includes a video decompression module 270, an inverse 3D motion compensation and inverse delta prediction module 272, a point cloud generation module 274, and a smoothing filter 276. In some embodiments, a decoder, such as decoder 280 may be combined with a decoder, such as decoder 230, or may share some components with the decoder, such as a video decompression module and/or smoothing filter. In decoder 280, the video/image streams are first decoded, then an inverse motion compensation and delta prediction procedure may be applied. The obtained images are then used in order to reconstruct a point cloud, which may be smoothed as described previously to generate a reconstructed point cloud 282.

Segmentation Process

[0098] FIG. 3A illustrates an example segmentation process for determining patches for a point cloud, according to some embodiments. The segmentation process as described in FIG. 3A may be performed by a decomposition into patches module, such as decomposition into patches module 206. A segmentation process may decompose a point cloud into a minimum number of patches (e.g., a contiguous subset of the surface described by the point cloud), while making sure that the respective patches may be represented by a depth field with respect to a patch plane. This may be done without a significant loss of shape information.

[0099] In some embodiments, a segmentation process comprises: [0100] Letting point cloud PC be the input point cloud to be partitioned into patches and {P(0), P(1) … , P(N-1)} be the positions of points of point cloud PC. [0101] In some embodiments, a fixed set D={D(0), D(1), … , D(K-1)} of K 3D orientations is pre-defined. For instance, D may be chosen as follows D={(1.0, 0.0, 0.0), (0.0, 1.0, 0.0), (0.0, 0.0, 1.0), (-1.0, 0.0, 0.0), (0.0, -1.0, 0.0), (0.0, 0.0, -1.0)} [0102] In some embodiments, the normal vector to the surface at every point P(i) is estimated. Any suitable algorithm may be used to determine the normal vector to the surface. For instance, a technique could include fetching the set H of the “N” nearest points of P(i), and fitting a plane H(i) to H(i) by using principal component analysis techniques. The normal to P(i) may be estimated by taking the normal V(i) to H(i). Note that “N” may be a user-defined parameter or may be found by applying an optimization procedure. “N” may also be fixed or adaptive. The normal values may then be oriented consistently by using a minimum-spanning tree approach. [0103] Normal-based Segmentation: An initial segmentation S0 of the points of point cloud PC may be obtained by associating respective points with the direction D(k) which maximizes the score (.gradient.(i)|D(k)), where <.|.> is the canonical dot product of R3. Pseudo code is provided below.

TABLE-US-00001 [0103] for (i = 0; i < pointCount; ++i) { clusterIndex = 0; bestScore = .gradient.(i)|D(0) ; for(j = 1; j < K; ++j) { score = .gradient.(i)|D(j) ; if (score > bestScore) { bestScore = score; clusterIndex = j; } } partition[i] = clusterIndex; }

[0104] Iterative segmentation refinement: Note that segmentation S0 associates respective points with the plane .PI.(i) that best preserves the geometry of its neighborhood (e.g. the neighborhood of the segment). In some circumstances, segmentation S0 may generate too many small connected components with irregular boundaries, which may result in poor compression performance. In order to avoid such issues, the following iterative segmentation refinement procedure may be applied: [0105] 1. An adjacency graph A may be built by associating a vertex V(i) to respective points P(i) of point cloud PC and by adding R edges {E(i,j(0)), … , E(i,j(R-1)} connecting vertex V(i) to its nearest neighbors {V(j(0)), V(j(1)), … , V(j(R-1))}. More precisely, {V(j(0)), V(j(1)), … , V(j(R-1))} may be the vertices associated with the points {P(j(0)), P(j(1)), … , P(j(R-1))}, which may be the nearest neighbors of P(i). Note that R may be a user-defined parameter or may be found by applying an optimization procedure. It may also be fixed or adaptive. [0106] 2. At each iteration, the points of point cloud PC may be traversed and every vertex may be associated with the direction D (k)* that maximizes*

[0106] ( .gradient. ( i ) D ( k ) + .lamda. R .zeta. ( i ) ) , ##EQU00001## where |.zeta.(i)| is the number of the R-nearest neighbors of V(i) belonging to the same cluster and .lamda. is a parameter controlling the regularity of the produced patches. Note that the parameters .lamda. and R may be defined by the user or may be determined by applying an optimization procedure. They may also be fixed or adaptive. In some embodiments, a “user” as referred to herein may be an engineer who configured a point cloud compression technique as described herein to one or more applications. [0107] 3.* An example of pseudo code is provided below*

TABLE-US-00002 [0107] for(I = 0; I < iterationCount; ++I) { for(i = 0; i < pointCount; ++i) { clusterIndex = partition[i]; bestScore = 0.0; for(k = 0; k < K; ++k) { score = .gradient.(i)|D(k) ; for(j .di-elect cons. {j(0), j(1), … , j(R – 1)}) { if (k == partition[j]) { score += .lamda. R ; ##EQU00002## } } if (score > bestScore) { bestScore = score; clusterIndex = k; } } partition[i] = clusterIndex; } }

[0108] In some embodiments, the pseudo code shown above may further include an early termination step. For example, if a score that is a particular value is reached, or if a difference between a score that is reached and a best score only changes by a certain amount or less, the search could be terminated early. Also, the search could be terminated if after a certain number of iterations (l=m), the clusterindex does not change. [0109] Patch segmentation: In some embodiments, the patch segmentation procedure further segments the clusters detected in the previous steps into patches, which may be represented with a depth field with respect to a projection plane. The approach proceeds as follows, according to some embodiments: [0110] 1. First, a cluster-based adjacency graph with a number of neighbors R’ is built, while considering as neighbors only the points that belong to the same cluster. Note that R’ may be different from the number of neighbors R used in the previous steps. [0111] 2. Next, the different connected components of the cluster-based adjacency graph are extracted. Only connected components with a number of points higher than a parameter .alpha. are considered. Let CC={CC(0), CC(1), … , CC(M-1)} be the set of the extracted connected components. [0112] 3. Respective connected component CC(m) inherits the orientation D(m) of the cluster it belongs to. The points of CC(m) are then projected on a projection plane having as normal the orientation D(m), while updating a depth map, which records for every pixel the depth of the nearest point to the projection plane. [0113] 4. An approximated version of CC(m), denoted C’(m), is then built by associating respective updated pixels of the depth map with a 3D point having the same depth. Let PC’ be the point cloud obtained by the union of reconstructed connected components {CC’(0), CC’(1), … , CC’(M-1)} [0114] 5. Note that the projection reconstruction process may be lossy and some points may be missing. In order, to detect such points, every point P(i) of point cloud PC may be checked to make sure it is within a distance lower than a parameter .delta. from a point of PC’. If this is not the case, then P(i) may be marked as a missed point and added to a set of missed points denoted MP. [0115] 6. The steps 2-5 are then applied to the missed points MP. The process is repeated until MP is empty or CC is empty. Note that the parameters .delta. and a may be defined by the user or may be determined by applying an optimization procedure. They may also be fixed or adaptive. [0116] 7. A filtering procedure may be applied to the detected patches in order to make them better suited for compression. Example filter procedures may include: [0117] a. A smoothing filter based on the geometry/texture/attributes of the points of the patches (e.g., median filtering), which takes into account both spatial and temporal aspects. [0118] b. Discarding small and isolated patches. [0119] c. User-guided filtering. [0120] d. Other suitable smoothing filter techniques.

Layers

[0121] The image generation process described above consists of projecting the points belonging to each patch onto its associated projection plane to generate a patch image. This process could be generalized to handle the situation where multiple points are projected onto the same pixel as follows: [0122] Let H(u,v) be the set of points of the current patch that get projected to the same pixel (u,v). Note that H(u,v) may be empty, may have one point or multiple points. [0123] If H(u,v) is empty then the pixel is marked as unoccupied. [0124] If the H(u,v) has a single element, then the pixel is filled with the associated geometry/texture/attribute value. [0125] If H(u,v), has multiple elements, then different strategies are possible: [0126] Keep only the nearest point P0(u,v) for the pixel (u,v) [0127] Take the average or a linear combination of a group of points that are within a distance d from P0(u,v), where d is a user-defined parameter needed only on the encoder side. [0128] Store two images: one for P0(u,v) and one to store the furthest point P1(u,v) of H(u,v) that is within a distance d from P0(u,v) [0129] Store N patch images containing a subset of H(u,v)

[0130] The generated patch images for point clouds with points at the same patch location, but different depths may be referred to as layers herein. In some embodiments, scaling/up-sampling/down-sampling could be applied to the produced patch images/layers in order to control the number of points in the reconstructed point cloud.

[0131] Guided up-sampling strategies may be performed on the layers that were down-sampled given the full resolution image from another “primary” layer that was not down-sampled.

[0132] In some embodiments, a delta prediction between layers could be adaptively applied based on a rate-distortion optimization. This choice may be explicitly signaled in the bit stream.

[0133] In some embodiments, the generated layers may be encoded with different precisions. The precision of each layer may be adaptively controlled by using a shift+scale or a more general linear or non-linear transformation.

[0134] In some embodiments, an encoder may make decisions on a scaling strategy and parameters, which are explicitly encoded in the bit stream. The decoder may read the information from the bit stream and apply the right scaling process with the parameters signaled by the encoder.

[0135] In some embodiments, a video encoding motion estimation process may be guided by providing a motion vector map to the video encoder indicating for each block of the image frame, a 2D search center or motion vector candidates for the refinement search. Such information, may be trivial to compute since the mapping between the 3D frames and the 2D image frames is available to the point cloud encoder and a coarse mapping between the 2D image frames could be computed by using a nearest neighbor search in 3D.

[0136] The video motion estimation/mode decision/intra-prediction could be accelerated/improved by providing a search center map, which may provide guidance on where to search and which modes to choose from for each N.times.N pixel block.

[0137] Hidden/non-displayed pictures could be used in codecs such as AV1 and HEVC. In particular, synthesized patches could be created and encoded (but not displayed) in order to improve prediction efficiency. This could be achieved by re-using a subset of the padded pixels to store synthesized patches.

[0138] The patch re-sampling (e.g., packing and patch segmentation) process described above exploits solely the geometry information. A more comprehensive approach may take into account the distortions in terms of geometry, texture, and other attributes and may improve the quality of the re-sampled point clouds.

[0139] Instead of first deriving the geometry image and optimizing the texture image given said geometry, a joint optimization of geometry and texture could be performed. For example, the geometry patches could be selected in a manner that results in minimum distortion for both geometry and texture. This could be done by immediately associating each possible geometry patch with its corresponding texture patch and computing their corresponding distortion information. Rate-distortion optimization could also be considered if the target compression ratio is known.

[0140] In some embodiments, a point cloud resampling process described above may additionally consider texture and attributes information, instead of relying only on geometry.

[0141] Also, a projection-based transformation that maps 3D points to 2D pixels could be generalized to support arbitrary 3D to 2D mapping as follows: [0142] Store the 3D to 2D transform parameters or the pixel coordinates associated with each point [0143] Store X, Y,* Z coordinates in the geometry images instead of or in addition to the depth information*

Packing

[0144] In some embodiments, depth maps associated with patches, also referred to herein as “depth patch images” or “geometry patch images,” such as those described above, may be packed into a 2D image frame. For example, a packing module, such as packing module 208, may pack depth patch images generated by a spatial image generation module, such as spatial image generation module 210. The depth maps, or depth patch images, may be packed such that (A) no non-overlapping block of T.times.T pixels contains depth information from two different patches and such that (B) a size of the generated image frame is minimized.

[0145] In some embodiments, packing comprises the following steps: [0146] a. The patches are sorted by height and then by width. The patches are then inserted in image frame (I) one after the other in that order. At each step, the pixels of image frame (I) are traversed in raster order, while checking if the current patch could be inserted under the two conditions (A) and (B) described above. If it is not possible then the height of (I) is doubled. [0147] b. This process is iterated until all the patches are inserted.

[0148] In some embodiments, the packing process described above may be applied to pack a subset of the patches inside multiples tiles of an image frame or multiple image frames. This may allow patches with similar/close orientations based on visibility according to the rendering camera position to be stored in the same image frame/tile, to enable view-dependent streaming and/or decoding. This may also allow parallel encoding/decoding.

[0149] In some embodiments, the packing process can be considered a bin-packing problem and a first decreasing strategy as described above may be applied to solve the bin-packing problem. In other embodiments, other methods such as the modified first fit decreasing (MFFD) strategy may be applied in the packing process.

[0150] In some embodiments, if temporal prediction is used, such as described for inter compression encoder 250, such an optimization may be performed with temporal prediction/encoding in addition to spatial prediction/encoding. Such consideration may be made for the entire video sequence or per group of pictures (GOP). In the latter case additional constraints may be specified. For example, a constraint may be that the resolution of the image frames should not exceed a threshold amount. In some embodiments, additional temporal constraints may be imposed, even if temporal prediction is not used, for example such as that a patch corresponding to a particular object view is not moved more than x number of pixels from previous instantiations.

[0151] FIG. 3B illustrates an example image frame comprising packed patch images and padded portions, according to some embodiments. Image frame 300 includes patch images 302 packed into image frame 300 and also includes padding 304 in space of image frame 300 not occupied by patch images. In some embodiments, padding, such as padding 304, may be determined so as to minimize incongruences between a patch image and the padding. For example, in some embodiments, padding may construct new pixel blocks that are replicas of, or are to some degree similar to, pixel blocks that are on the edges of patch images. Because an image and/or video encoder may encode based on differences between adjacent pixels, such an approach may reduce the number of bytes required to encode an image frame comprising of patch images and padding, in some embodiments.

[0152] In some embodiments, the patch information may be stored in the same order as the order used during the packing, which makes it possible to handle overlapping 2D bounding boxes of patches. Thus a decoder receiving the patch information can extract patch images from the image frame in the same order in which the patch images were packed into the image frame. Also, because the order is known by the decoder, the decoder can resolve patch image bounding boxes that overlap.

[0153] FIG. 3C illustrates an example image frame 312 with overlapping patches, according to some embodiments. FIG. 3C shows an example with two patches (patch image 1 and patch image 2) having overlapping 2D bounding boxes 314 and 316 that overlap at area 318. In order to determine to which patch the T.times.T blocks in the area 318 belong, the order of the patches may be considered. For example, the T.times.T block 314 may belong to the last decoded patch. This may be because in the case of an overlapping patch, a later placed patch is placed such that it overlaps with a previously placed patch. By knowing the placement order it can be resolved that areas of overlapping bounding boxes go with the latest placed patch. In some embodiments, the patch information is predicted and encoded (e.g., with an entropy/arithmetic encoder). Also, in some embodiments, U0, V0, DU0 and DV0 are encoded as multiples of T, where T is the block size used during the padding phase.

[0154] FIG. 3C also illustrates blocks of an image frame 312, wherein the blocks may be further divided into sub-blocks. For example block A1, B1, C1, A2, etc. may be divided into multiple sub-blocks, and, in some embodiments, the sub-blocks may be further divided into smaller blocks. In some embodiments, a video compression module of an encoder, such as video compression module 218 or video compression module 264, may determine whether a block comprises active pixels, non-active pixels, or a mix of active and non-active pixels. The video compression module may budget fewer resources to compress blocks comprising non-active pixels than an amount of resources that are budgeted for encoding blocks comprising active pixels. In some embodiments, active pixels may be pixels that include data for a patch image and non-active pixels may be pixels that include padding. In some embodiments, a video compression module may sub-divide blocks comprising both active and non-active pixels, and budget resources based on whether sub-blocks of the blocks comprise active or non-active pixels. For example, blocks A1, B1, C1, A2 may comprise non-active pixels. As another example block E3 may comprise active pixels, and block B6, as an example, may include a mix of active and non-active pixels.

[0155] In some embodiments, a patch image may be determined based on projections, such as projecting a point cloud onto a cube, cylinder, sphere, etc. In some embodiments, a patch image may comprise a projection that occupies a full image frame without padding. For example, in a cubic projection each of the six cubic faces may be a patch image that occupies a full image frame.

[0156] For example, FIG. 3D illustrates a point cloud being projected onto multiple projections, according to some embodiments.

[0157] In some embodiments, a representation of a point cloud is encoded using multiple projections. For example, instead of determining patches for a segment of the point cloud projected on a plane perpendicular to a normal to the segment, the point cloud may be projected onto multiple arbitrary planes or surfaces. For example, a point cloud may be projected onto the sides of a cube, cylinder, sphere, etc. Also multiple projections intersecting a point cloud may be used. In some embodiments, the projections may be encoded using conventional video compression methods, such as via a video compression module 218 or video compression module 264. In particular, the point cloud representation may be first projected onto a shape, such as a cube, and the different projections/faces projected onto that shape (i.e. front (320), back (322), top (324), bottom (326), left (328), right (330)) may all be packed onto a single image frame or multiple image frames. This information, as well as depth information may be encoded separately or with coding tools such as the ones provided in the 3D extension of the HEVC (3D-HEVC) standard. The information may provide a representation of the point cloud since the projection images can provide the (x,y) geometry coordinates of all projected points of the point cloud. Additionally, depth information that provides the z coordinates may be encoded. In some embodiments, the depth information may be determined by comparing different ones of the projections, slicing through the point cloud at different depths. When projecting a point cloud onto a cube, the projections might not cover all point cloud points, e.g. due to occlusions. Therefore additional information may be encoded to provide for these missing points and updates may be provided for the missing points.

[0158] In some embodiments, adjustments to a cubic projection can be performed that further improve upon such projections. For example, adjustments may be applied at the encoder only (non-normative) or applied to both the encoder and the decoder (normative).

[0159] More specifically, in some embodiments alternative projections may be used. For example, instead of using a cubic projection, a cylindrical or spherical type of a projection method may be used. Such methods may reduce, if not eliminate, redundancies that may exist in the cubic projection and reduce the number or the effect of “seams” that may exist in cubic projections. Such seams may create artifacts at object boundaries, for example. Eliminating or reducing the number or effect of such seams may result in improved compression/subjective quality as compared to cubic projection methods. For a spherical projection case, a variety of sub-projections may be used, such as the equirectangular, equiangular, and authagraph projection among others. These projections may permit the projection of a sphere onto a 2D plane. In some embodiments, the effects of seams may be de-emphasized by overlapping projections, wherein multiple projections are made of a point cloud, and the projections overlap with one another at the edges, such that there is overlapping information at the seams. A blending effect could be employed at the overlapping seams to reduce the effects of the seams, thus making them less visible.

[0160] In addition to, or instead of, considering a different projection method (such as cylindrical or spherical projections), in some embodiments multiple parallel projections may be used. The multiple parallel projections may provide additional information and may reduce a number of occluded points. The projections may be known at the decoder or signaled to the decoder. Such projections may be defined on planes or surfaces that are at different distances from a point cloud object. Also, in some embodiments the projections may be of different shapes, and may also overlap or cross through the point cloud object itself. These projections may permit capturing some characteristics of a point cloud object that may have been occluded through a single projection method or a patch segmentation method as described above.

[0161] For example, FIG. 3E illustrates a point cloud being projected onto multiple parallel projections, according to some embodiments. Point cloud 350 which includes points representing a coffee mug is projected onto parallel horizontal projections 352 that comprise planes orthogonal to the Z axis. Point cloud 350 is also projected onto vertical projections 354 that comprise planes orthogonal to the X axis, and is projected onto vertical projections 356 that comprise planes orthogonal to the Y axis. In some embodiments, instead of planes, multiple projections may comprise projections having other shapes, such as multiple cylinders or spheres.

Generating Images Having Depth

[0162] In some embodiments, only a subset of the pixels of an image frame will be occupied and may correspond to a subset of 3D points of a point cloud. Mapping of patch images may be used to generate geometry, texture, and attribute images, by storing for each occupied pixel the depth/texture/attribute value of its associated point.

[0163] In some embodiments, spatial information may be stored with various variations, for example spatial information may: [0164] a. Store depth as a monochrome image. [0165] b. Store depth as Y and keep U and V empty (where YUV is a color space, also RGB color space may be used). [0166] c. Store depth information for different patches in different color planes Y, U and V, in order to avoid inter-patch contamination during compression and/or improve compression efficiency (e.g., have correlated patches in the same color plane). Also, hardware codec capabilities may be utilized, which may spend the same encoding\decoding time independently of the content of the frame. [0167] d. Store depth patch images on multiple images or tiles that could be encoded and decoded in parallel. One advantage is to store depth patch images with similar/close orientations or based on visibility according to the rendering camera position in the same image/tile, to enable view-dependent streaming and/or decoding. [0168] e. Store depth as Y and store a redundant version of depth in U and V. [0169] f. Store X, Y, Z coordinates in Y, U, and V [0170] g. Different bit depth (e.g., 8, 10 or 12-bit) and sampling (e.g., 420, 422, 444 … ) may be used. Note that different bit depth may be used for the different color planes.

Padding

[0171] In some embodiments, padding may be performed to fill the non-occupied pixels with values such that the resulting image is suited for video/image compression. For example, image frame padding module 216 or image padding module 262 may perform padding as described below.

[0172] In some embodiments, padding is applied on pixels blocks, while favoring the intra-prediction modes used by existing video codecs. More precisely, for each block of size B.times.B to be padded, the intra prediction modes available at the video encoder side are assessed and the one that produces the lowest prediction errors on the occupied pixels is retained. This may take advantage of the fact that video/image codecs commonly operate on pixel blocks with pre-defined sizes (e.g., 64.times.64, 32.times.32, 16.times.16 … ). In some embodiments, other padding techniques may include linear extrapolation, in-painting techniques, or other suitable techniques.

Video Compression

[0173] In some embodiments, a video compression module, such as video compression module 218 or video compression module 264, may perform video compression as described below.

[0174] In some embodiments, a video encoder may leverage an occupancy map, which describes for each pixel of an image whether it stores information belonging to the point cloud or padded pixels. In some embodiments, such information may permit enabling various features adaptively, such as de-blocking, adaptive loop filtering (ALF), or shape adaptive offset (SAO) filtering. Also, such information may allow a rate control module to adapt and assign different, e.g. lower, quantization parameters (QPs), and in an essence a different amount of bits, to the blocks containing the occupancy map edges. Coding parameters, such as lagrangian multipliers, quantization thresholding, quantization matrices, etc. may also be adjusted according to the characteristics of the point cloud projected blocks. In some embodiments, such information may also enable rate distortion optimization (RDO) and rate control/allocation to leverage the occupancy map to consider distortions based on non-padded pixels. In a more general form, weighting of distortion may be based on the “importance” of each pixel to the point cloud geometry. Importance may be based on a variety of aspects, e.g. on proximity to other point cloud samples, directionality/orientation/position of the samples, etc. Facing forward samples, for example, may receive a higher weighting in the distortion computation than backward facing samples. Distortion may be computed using metrics such as Mean Square or Absolute Error, but different distortion metrics may also be considered, such as SSIM, VQM, VDP, Hausdorff distance, and others.

Occupancy Map Compression

[0175] In some embodiments, an occupancy map compression module, such as occupancy map compression module 220, may compress an occupancy map as described below.

Example Occupancy Map Compression Techniques

[0176] In some embodiments, an occupancy map may be encoded in a hierarchical mode. Such a process may comprise: [0177] 1. A binary information for each B1.times.B2 pixel block (e.g., a rectangle that covers the entire image frame, or smaller blocks of different sizes such as 64.times.64, 64.times.32, 32.times.32 block, etc.) being encoded indicating whether the block is empty (e.g., has only padded pixels) or non-empty (e.g., has non-padded pixels). [0178] 2. If the block is non-empty, then a second binary information may be encoded to indicate whether the block is full (e.g., all the pixels are non-padded) or not. [0179] 3. The non-empty and non-full blocks may then be refined by considering their (B1/2).times.(B2/2) sub-blocks. [0180] 4. The steps 1-3 may be repeated until the size of the block reaches a certain block size B3.times.B4 (e.g., of size 4.times.4). At this level only the empty/non-empty information may be encoded. [0181] 5. An entropy-based codec may be used to encode the binary information in steps 1 and 2. For instance, context adaptive binary arithmetic encoders may be used. [0182] 6. The reconstructed geometry image may be leveraged to better encode the occupancy map. More precisely, the residual prediction errors may be used to predict whether a block is empty or not or full or not. Such an information may be incorporated by using a different context based on the predicted case or simply by encoding the binary value XORed with the predicted value.

[0183] In some embodiments, mesh-based codecs may be an alternative to the approach described above.

Additional Example Occupancy Map Compression Technique

[0184] In some embodiments, auxiliary information and the patch encoding order may be leveraged in order to efficiently compress a mapping information indicating for each T.times.T block of an image frame (e.g., 16.times.16 block) to which patch it belongs to. This mapping may be explicitly encoded in the bit stream as follows: [0185] A list of candidate patches is created for each T.times.T block of an image frame by considering all the patches that overlap with that block. [0186] The list of candidates is sorted in the reverse order of the patches. Meaning the candidate patches are ordered from smallest patch to largest patch. [0187] For each block, the index of the patches in this list is encoded by using an arithmetic or other form of an entropy encoder (e.g. UVLC or Huffman based). [0188] Note that empty blocks are assigned a special index, such as zero. [0189] The mapping information described above makes it possible to detect empty T.times.T blocks of an image frame (e.g., blocks that contain only padded pixels). The occupancy information is encoded only for the non-empty T.times.T blocks (e.g., the blocks that contain at least one non-padded pixel). [0190] The occupancy map is encoded with a precision of a B0.times.B0 blocks. In order to achieve lossless encoding B0 is chosen to be 1. In some embodiments B0=2 or B0=4, which may result in visually acceptable results, while significantly reducing the number of bits required to encode the occupancy map. [0191] Binary values are associated with B0.times.B0 sub-blocks belonging to the same T.times.T block. Different strategies are possible. For instance, one could associate a value of 1 if the sub-block contains at least some non-padded pixels and 0 otherwise. If a sub-block has a value of 1 it is said to be full, otherwise it is an empty sub-block. [0192] If all the sub-blocks of a T.times.T block are full (e.g., have value 1). The block is said to be full. Otherwise, the block is said to be non-full. [0193] A binary information is encoded for each T.times.T block to indicate whether it is full or not. Various encoding strategies could be used. For instance, a context adaptive binary arithmetic encoder could be used. [0194] If the block is non-full, an extra information is encoded indicating the location of the full/empty sub-blocks. More precisely, the process may proceed as follows: [0195] Different traversal orders are defined for the sub-blocks. FIG. 12B, shows some examples. The traversal orders are predetermined and known to both the encoder and decoder. [0196] The encoder chooses one of the traversal orders and explicitly signals its index in the bit stream. [0197] The binary values associated with the sub-blocks are encoded by using a run-length encoding strategy. [0198] The binary value of the initial sub-block is encoded. Various encoding strategies could be used. For instance, fixed length coding or a context adaptive binary arithmetic encoder could be used. [0199] Continuous runs of 0s and 1s are detected, while following the traversal order selected by the encoder. [0200] The number of detected runs is encoded. Various encoding strategies could be used. For instance, fixed length coding or a context adaptive binary arithmetic encoder, or a universal variable length encoder (UVLC) could be used. [0201] The length of each run, except of the last one, is then encoded. Various encoding strategies could be used. For instance, fixed length coding, a context adaptive binary arithmetic encoder, or a universal variable length encoder could be used.

[0202] Note that the symbol probabilities used during the arithmetic encoding could be initialized by using values explicitly signaled in the bit stream by the encoder in order to improve compression efficiency. Such information could be signaled at frame, slice, row(s) of blocks, or block level, or using a non-fixed interval. In that case, a system may have the ability to signal the initialization interval, or the interval adaptation could be predefined between encoder and decoder. For example, the interval could start with one block, and then increment by one block afterwards (e.g. using an adaptation positions of {1, 2, 3 … N-1 … } blocks.

[0203] The choice of the traversal order may have a direct impact on the compression efficiency. Different strategies are possible. For instance, the encoder could choose the traversal order, which would result in the lowest number of bits or the lowest number of runs. In some embodiments, hierarchical sub-blocks with variable sizes may be used.

[0204] In some embodiments, temporal prediction may be used for encoding/compressing occupancy maps as follows: [0205] a. The occupancy map of the current frame may be predicted from the occupancy map of a reference frame (e.g. through a difference process assuming zero motion). The prediction could be done at the frame level, but could also be done at a sub-block level, e.g. signal 1 bit whether a block will be predicted temporally, or the original map for a block will be used instead. [0206] b. Prediction could be enhanced by using motion compensation and by associating a motion vector with each T.times.T block. [0207] c. The values of the current block may be XOR-ed with the values of the block referenced by the motion vector or the co-located block. If no prediction is used, the current block may be coded as is. [0208] d. Motion vectors could be integer, integer multiples, or can have sub-pixel precision. [0209] e. The encoding strategy described above may be applied to the results. [0210] f. The motion vectors of the current block may be predicted based on the motion vectors of the previously encoded blocks. For example, a list of candidate predicted motion vectors may be computed based on the motion vectors of spatially and/or temporally neighboring blocks that have already been encoded. The index of the best candidate to be used as a predictor and the difference can be explicitly encoded in the bit stream. The process may be similar to the process used in codecs such as AVC and HEVC among others. A reduction in temporal candidates may be performed similar to what is done in HEVC to reduce memory requirements. The residual motion vector can then be encoded using a technique such as context adaptive arithmetic encoding or UVLC. [0211] g. A skip mode may also be supported to indicate that the predicted block matches exactly the reference block. In that case, no residual motion vector is needed. [0212] h. Different block sizes could be used instead of sticking with T.times.T blocks. [0213] i. The choice of the block size and the motion vectors could be achieved by minimizing the number of bits required to encode the occupancy map. [0214] j. The process could also consider multiple references.

[0215] In some embodiments, additional techniques for encoding/compression of an occupancy map may include: [0216] Using clues included in the video picture to help encode the occupancy map, such as: [0217] Use high quantization parameters QPs (e.g., 51) or use skip mode for blocks composed of padded pixels only. [0218] The arithmetic encoding contexts could be adaptively adjusted based on information extracted from the video bit streams associated with the texture/geometry/motion frames. [0219] Group the binary values associated with pixels into 8-bit or 10-bit words and encode them with dictionary-based approaches such as the DEFLATE algorithm. [0220] Pixels could be grouped 4.times.2/5.times.2 blocks or by leveraging a zig zag scan. [0221] Only the pixels belonging to non-empty T.times.T blocks may be encoded. [0222] The mapping information indicating for each T.times.T block to which patch it belongs may encoded.

Additional Example Occupancy Map Compression Techniques

[0223] In some embodiments, a binary occupancy map is generated based on whether or not bocks of the occupancy map are occupied or un-occupied. This may be performed in a similar manner as described above. Also, the patch information (e.g. bounding box position, size, etc.) is encoded using an arithmetic encoder, in a similar manner as described above. However, instead of relying on the occupancy map to discard empty blocks that intersect with at least one patch bounding box, the empty boxes are explicitly signaled with a special value for the local index. In this approach, the block to patch information is decoded when needed.

[0224] In some embodiments, instead of using an arithmetic encoder as described above to encode block to patch information that links boxes of the occupancy map with particular patches, the block to patch information (which contains the local indexes) may be encoded using a video-based encoder. The encoded block-to patch information may then be decoded using a corresponding video-decoder.

[0225] In some embodiments, instead of generating a binary occupancy map based on whether or not bocks of the occupancy map are occupied or un-occupied, a non-binary occupancy map is generated. The non-binary occupancy map is configured such that each pixel not only indicates whether the pixel is occupied or non-occupied, but also includes an attribute value, such as a color value that is associated with a local index value of a patch with which the pixel is associated. If the pixel is non-occupied, the pixel may have a color value of zero. Also, the patch information (e.g. bounding box position, size, etc.) is encoded using an arithmetic encoder, in a similar manner as described above. The non-binary occupancy map may be encoded using a video-based encoder. A decoder can retrieve the block to patch information by decoding the non-binary occupancy map and matching each pixel value with the local index lists.

[0226] In some embodiments, instead of using a local index, a full list of patches may be used as an index. In such embodiments, there may be no need to compute a list of candidate patches for each block. The decoder can retrieve the block-to-patch information by decoding the non-binary occupancy map directly reading the index value for the patch associated with the pixel from the value of the pixel. In such embodiments, the local index may be omitted because there are enough unique values (e.g. non-binary) values available to be associated with a block, such that each candidate patch may be assigned a unique value.

[0227] In some embodiments, during the generation of the occupancy map, the bounding boxes for the patches may be adjusted or initially packed in an image frame such that the bounding boxes do not overlap. This removes ambiguity as to whether a particular bounding box belongs to a particular patch or another patch. The patch information (with non-overlapping bounding boxes) is encoding using an arithmetic encoder. Because there is not ambiguity as to which patch goes with which bounding box, the block to patch information (such as in the local index or complete index, as described above), may be omitted.

[0228] In some embodiments, a process that uses a full list of patches (instead of a local index) may result in a high number of patches, which may exceed the max possible number of values (e.g. color values) that may be represented in the non-binary occupancy map. In some embodiments, to address such issues, an occupancy map may be decomposed into segments, with a limited number of patches per segments. Thus for each segment, the patch index is bound. For example, fewer patches may be listed as possibilities for a segment of an occupancy map, such that for each segment the list of possible patches is less than the max possible number of values (e.g. color values). In some such embodiments, bounding boxes for different patches may be allowed to overlap within a segment, but not across segments. During decoding, each segment may have its own global index list of possible patches for that segment.

[0229] In some embodiments, a binary occupancy map is generated such that when the patches are packed in the image frame, a bounding box of the patch, aligned to an occupancy resolution does not intersect any previously packed patches of size=_occupancy resolution*size occupancy resolution (e.g. a 16.times.16 block). The patch information (e.g. bounding box position and size) for each patch is encoded using an arithmetic encoder. The order in which the patch information for each patch is encoded may create a hierarchy of patches, such that for any overlapping bounding boxes, the corresponding patch that goes with the bound box can be resolved based on the hierarchy of patch information. The decoder may reconstruct block to patch information using the arithmetically encoded patch information (without the block to patch information being explicitly encoded). For example, a patch list may be parsed in a same order at a decoder as an order in which the patch list was generated at encoding time, wherein the order indicates an order in which the patches were packed in the image frame. This is possible because the packing guarantees that the bounding box for a given patch does not cover any previously processed patch. In such embodiments, patches may be packed (and signaled) in an order such as from small to large, or vice versa. During the packing, each block may include pixels of just one patch, but some bounding boxes for multiple patches may overlap, wherein blocks of the overlapping patches include no pixels for either patch, or pixels for just one of the patches, but not pixels for more than one patch.

Auxiliary Patch-Information Compression

[0230] In some embodiments, for each patch, the following information may be encoded. For example, by auxiliary patch-info compression module 222. [0231] Its location (U0, V0) in the packed image frame and the extent of its 2D bounding box (DU0, DV0). [0232] Minimum/maximum/average/median depth value. [0233] Index of the projection direction.

Video-Based Occupancy Map Compression

[0234] As described above, in some embodiments, the occupancy map is a binary information that indicates for each pixel in the image frame whether the pixel should be interpreted as an actual point in the point cloud or not, e.g. the pixel is a padded pixel or not. Also, as described above, the auxiliary patch-information indicates for each T.times.T block of an image frame to which patch it belongs. Whereas it was described above to encode an index of patches for a block of an image frame and to keep encoding information for sub-blocks of the image frame until the sub-blocks were either fully empty or fully occupied, an alternative approach is to use a video encoder to encode an additional image frame for the occupancy map. In such embodiments, the additional occupancy map image frame, indicates occupied and unoccupied pixels based on image properties such as different colors (e.g. occupied pixels may be white and non-occupied pixels may be black). In this way it is not necessary to completely subdivide the blocks of the image frame until only occupied or un-occupied sub-blocks are determined. Instead, it is only necessary to identify bounding box sizes and locations in the image frame for the respective patches. The video encoded occupancy map will mirror the image frame and the different pixel values in the occupancy map video image frame will indicate which pixels in a given bounding box of the patch video image frame are patch image pixels or are padded pixels. Thus there is not a need to create a bit stream of sub-divided blocks of the image frame and there is not a need to indicate for each sub-block whether the sub-block is full or empty. Instead the video encoded occupancy map can be used to determine which pixels included in a patch bounding box are padded pixels or patch image pixels. In some embodiments, an occupancy map may be first encoded and then used to generate an index of patches that are associated with blocks of an image frame. In some embodiments, a compression process follows the following procedure that leverages existing video codecs to compress an occupancy map.

[0235] The occupancy map could be encoded with a precision of B0.times.B1 blocks. In order to achieve lossless encoding B0 and B1 may be chosen to be equal to 1. In practice B0=B1=2 or B0=B1=4 may result in visually acceptable results, while significantly reducing the number of bits required to encode the occupancy map.

[0236] In some embodiments, a single binary is associated with each B0.times.B1 sub-block. Different strategies are possible. For instance, one could associate a value of 1 with the rule that the sub-block contains at least one non-padded pixel and the value of 0 if not. In order to reduce computational complexity, the binary values of multiple B0.times.B1 blocks could be grouped together in a single pixel value.

[0237] A binary video frame may be generated by storing the value of each B0.times.B1 block in a pixel. The obtained video frame could be compressed by using a lossless video codec. For example the HEVC video codec could be utilized and its main, screen context coding (scc) main or other profiles could be used.

[0238] In some embodiments, the occupancy map could be packed in a 4:4:4 or 4:2:0 chroma format, where the chroma information could contain fixed values, e.g. the values 0 or 128 for an 8 bit codec. The occupancy map could also be coded using a codec supporting a monochrome representation. The occupancy map could be replicated in all color components and encoded using a 4:4:4 representation. Other rearrangements of the occupancy map could be used so as to fit the data in a 4:4:4, 4:2:2, or 4:2:0 representation, while preserving the lossless nature of the signal and at the same time preserving the lossless characteristics of the occupancy map. For example, the occupancy map could be segmented to even horizontal and odd horizontal position sub-maps, and those sub-maps could be embedded into a 4:4:4 signal, the odd position samples in the Y plane and the even position samples in the U plane, and then encoded. This could provide savings in complexity since a reduced resolution (by half) image would be encoded. Other such arrangements could be used.

[0239] The occupancy map is used to detect non-empty T.times.T blocks and only for those blocks a patch index is encoded by proceeding as follows: [0240] 1) A list of candidate patches is created for each T.times.T block by considering all the patches that contain that block. [0241] 2) The list of candidates is sorted in the reverse order of the patches. Meaning the index is sorted from smallest patch to largest patch, e.g. the patches with bounding boxes covering the smallest area are ordered ahead of patches with bounding boxes covering larger areas of the patch image frame. [0242] 3) For each block, the index of the patch in this list is encoded by using an entropy encoder, e.g. an arithmetic encoder or other suitable encoder.

Patch Alignment and Size Determination in a 2D Bounding Box of an Occupancy Map

[0243] In some embodiments, methods may be applied to remove redundant output points created by the occupancy map quantization/downsampling/upsampling process. By removing these points, the reconstruction process can result in better reconstruction. Furthermore, fewer points may need to be processed during post-processing, e.g. when performing smoothing as described below, thus reducing reconstruction complexity as well as during attribute image generation during encoding. Additionally, quality of the “removed” points in the geometry and attribute layers may be less important, therefore the characteristics of such points may be exploited during compression, such as devoting fewer resources to redundant points that will be removed.

[0244] In some embodiments, when a patch is created, the patch size information (e.g. sizeU0, sizeV0) is defined as multiples of the occupancy packing block. In other words, when patch size is N.times.M and the occupancy packing block resolution is 16, sizeU0 and sizeV0 will be (16*(N/16+1), 16*(M/16+1)). For example, Table 1 shows an example algorithm for determining the width and width of a 2D bounding box for a patch.

TABLE-US-00003 TABLE 1 Width and Height of Patch Derivation If p is equal to 0, then: Patch2dSizeU[ frmIdx ][ p ] = pdu_2d_delta_size_u[ frmIdx ][ p ] * ops_occupancy_packing_block_size (8-8) Patch2dSizeV[ frmIdx ][ p ] = pdu_2d_delta_size_v[ frmIdx ][ p ] * ops_occupancy_packing_block_size (8-9) Otherwise, if (p > 0), then: Patch2dSizeU[ frmIdx ][ p ] = Patch2dSizeU[ frmIdx ][ p – 1 ] + pdu_2d_delta_size_u[ frmIdx ][ p ] * ops_occupancy_packing_block_size (8-10) Patch2dSizeV[ frmIdx ][ p ] = Patch2dSizeV[ frmIdx ][ p – 1 ] + pdu_2d_delta_size_v[ frmIdx ][ p ] * ops_occupancy_packing_block_size (8-11)

[0245] In some embodiments, in a patch bounding box, there could be “empty” lines and/or columns maximum equal to (occupancy packing block resolution -1).

[0246] In some embodiments, an occupancy map could be quantized/downsampled by oPrecision which can be derived from the decoded occupancy map video resolution and the nominal resolution of the decoded video frames and then dequantized/upsampled when it is used. Therefore, (oPrecision.times.oPrecision) pixels will share one same value (1. Occupied 0. Empty). When the (oPrecision.times.oPrecision) pixels were not fully filled with 1 before the quantization process, the dequantization process will mark previously empty pixels with redundant points, and it would add on the distortion and complexity of the point cloud.

[0247] A method which simply discards samples that would have otherwise created additional points may result in holes or crack during reconstruction of the point cloud. A method which moves occupied samples to reduce redundant pixels may, for irregular shapes, result in redundant pixels.

[0248] In some embodiments, to improve upon such methods and to remove redundant output points, the width, height, and placement of a patch in an occupancy map may be adjusted.

Point Cloud Resampling

[0249] In some embodiments, a point cloud resampling module, such as point cloud resampling module 252, may resample a point cloud as described below.

[0250] In some embodiments, dynamic point clouds may have a different number of points from one frame to another. Efficient temporal prediction may require mapping the points of the current frame, denoted CF, to the points of a reference frame, denoted RF. Signaling such a mapping in a bit stream may require a high number of bits and thus may be inefficient. Instead, re-sampling of a current frame CF may be performed so that the current frame CF has the same number of points as reference frame RF. More precisely, the points of reference frame RF may be displaced such that its shape matches the shape of current frame CF. As a second step, the color and attributes of current frame CF may be transferred to the deformed version of reference frame RF. The obtained frame CF’ may be considered as the re-sampled version of the current frame. The decision to compress the approximation CF’ of CF may be made by comparing the rate distortion costs of both options (e.g., encoding CF’ as inter-frame vs. encoding CF as intra-frame). In some embodiments, pre-adjusting RF may be performed in an effort to make it a better reference for future CF images. Resampling may comprise the following: [0251] a. First, normals of the points associated with current frame CF and reference frame RF may be estimated and oriented consistently. For every point P belonging to current frame CF (resp. Q belonging to RF), let .alpha.(P) (resp., .alpha.(Q)) be its position and .gradient.(P) (resp., .gradient.(Q)) its normal. A 6D vector, denoted .upsilon.(P) (resp., .upsilon.(Q)) is then associated with every point by combining its position and a weighted version of its normal in the same vector.

[0251] .upsilon. ( P ) = [ .alpha. ( P ) .gradient. ( P ) ] .upsilon. ( Q ) = [ .alpha. ( Q ) .gradient. ( Q ) ] , ##EQU00003## where .epsilon. is a parameter controlling the importance of normal for positions. .epsilon. could be defined by the user or could be determined by applying an optimization procedure. They could also be fixed of adaptive. [0252] b. Two mappings from reference frame RF to current frame CF and from current frame CF to reference frame RF are computed as follows: [0253] i. Every point Q of reference frame RF is mapped to the point P(Q) of current frame CF that has the minimum distance to Q in the 6D space defined in the previous step. [0254] ii. Every point P of current frame CF is mapped to the point Q(P) of reference frame RF that has the minimum distance to P in the 6D space defined in the previous step. Let .rho.(Q) be the set of points of current frame CF that are mapped to the same point Q. [0255] c. At each iteration [0256] i. The positions of the points of reference frame RF are updated as follows:

[0256] .alpha. ’ ( Q ) = w .alpha. ( P ( Q ) ) + ( 1 – w ) .rho. ( Q ) P .di-elect cons. .rho. ( Q ) .alpha. ( P ) , ##EQU00004## where |.rho.(Q)| is the number of elements of .rho.(Q). The parameter w could be defined by the user or could be determined by applying an optimization procedure. It could also be fixed or adaptive. [0257] ii. The previous updated step results usually in an irregular repartition of the points. In order to overcome such limitations, a Laplacian-based smoothing procedure is applied. The idea is to update the positions of the points such that they stay as close as possible to {.alpha.’(Q)}, while favoring a repartition as close as possible to the original point repartition in reference frame RF. More precisely, the following sparse linear system may be solved:

[0257] { .alpha. * ( Q ) } = argmin { .alpha. ’ ( Q ) } { .SIGMA. Q .di-elect cons. RF .alpha. ” ( Q ) – .alpha. ’ ( Q ) 2 + .gamma..SIGMA. Q .di-elect cons. RF .alpha. ” ( Q ) – 1 R .SIGMA. Q ’ .di-elect cons. N ( Q ) .alpha. ” ( Q ’ ) – .alpha. ( Q ) – 1 R .SIGMA. Q ’ .di-elect cons. N ( Q ) .alpha. ( Q ’ ) 2 } , ##EQU00005## where N(Q) is the set of the R nearest neighbors of Q in reference frame RF. [0258] iii. The mappings between the updated RF’ point cloud and current frame CF are then updated as follows [0259] 1. Every point Q of RF’ is mapped to the point P(Q) of current frame CF that has the minimum distance to Q in the 3D space of positions. [0260] 2. Every point P of current frame CF is mapped to the point Q(P) of reference frame RF that has the minimum distance to P in the 3D space of positions. Let .rho.(Q) be the set of points of current frame CF that are mapped to the same point Q. [0261] d. This process is iterated until a pre-defined number of iterations is reached or there is no further change. [0262] e. At this stage, the color and attribute information is transferred from current frame CF to RF’* by exploiting the following formula*

[0262] A ( Q ) = w ( A ) . A ( P ( Q ) ) + ( 1 – w ( A ) ) .rho. ( Q ) .SIGMA. P .di-elect cons. .rho. ( Q ) A ( P ) , ##EQU00006## where A stands for the texture or attribute to be transferred, |.rho.(Q)| is the number of elements of .rho.(Q). The parameter w(A) could be defined by the user or could be determined by applying an optimization procedure. It could also be fixed of adaptive.

3D Motion Compensation

[0263] In some embodiments, the positions, attributes and texture information may be temporally predicted by taking the difference between the value at current resampled frame minus a corresponding value, e.g. motion compensated value, from the reference frame. These values may be fed to the image generation stage to be stored as images. For example, such techniques may be performed by 3D motion compensation and delta vector prediction module 254.

Smoothing Filter

[0264] In some embodiments, a smoothing filter of a decoder, such as smoothing filter 244 or smoothing filter 276 of decoder 230 or decoder 280, may perform smoothing as described below.

[0265] In some embodiments, a reconstructed point cloud may exhibit discontinuities at the patch boundaries, especially at very low bitrates. In order to alleviate such a problem, a smoothing filter may be applied to the reconstructed point cloud. Applying the smoothing filter may comprise: [0266] a. By exploiting the occupancy map, both the encoder and the decoder may be able to detect boundary points, which are defined as being points belonging to B0.times.B0 blocks encoded during the last iteration of the hierarchical occupancy map compression procedure described in previous sections above. [0267] b. The boundary points may have their positions/attribute/texture updated. More precisely, respective boundary points may be assigned a smoothed position based on its R nearest neighbors in the point cloud. The smoothed position may be the centroid/median of the nearest neighbors. Another option may comprise fitting a plane or any smooth surface the nearest neighbor and assigning as a smoothed position the projection of the point on that surface. The number of parameters and/or the smoothing strategy may be chosen by a user or determined by applying an optimization strategy. They may be fixed for all the points or chosen adaptively. These parameters may be signaled in the bit stream. [0268] c. In order to reduce the computational complexity of the smoothing stage, a subsampled version of the reconstructed point cloud may be considered when looking for the nearest neighbors. Such subsampled version could be efficiently derived by considering a subsampled version of the geometry image and the occupancy map.

Closed-Loop Color Conversion

[0269] In some embodiments, an encoder and/or decoder for a point cloud may further include a color conversion module to convert color attributes of a point cloud from a first color space to a second color space. In some embodiments, color attribute information for a point cloud may be more efficiently compressed when converted to a second color space. For example, FIGS. 4A and 4B illustrates encoders 400 and 450 which are similar encoders as illustrated in FIGS. 2A and 2C, but that further include color conversion modules 402 and 404, respectively. While not illustrated, decoders such as the decoders illustrated in FIGS. 2B and 2D, may further include color conversion modules to convert color attributes of a decompressed point cloud back into an original color space, in some embodiments.

[0270] FIG. 4C illustrates components of a closed-loop color conversion module, according to some embodiments. The closed-loop color conversion module 410 illustrated in FIG. 4C may be a similar closed-loop color conversion module as closed-loop color conversion modules 402 and 404 illustrated in FIGS. 4A and 4B.

[0271] In some embodiments, a closed-loop color conversion module, such as closed-loop color conversion module 410, receives a compressed point cloud from a video encoder, such as video compression module 218 illustrated in FIG. 4A or video compression module 264 illustrated in FIG. 4B. Additionally, a closed-loop color conversion module, such as closed-loop color conversion module 410, may receive attribute information about an original non-compressed point cloud, such as color values of points of the point cloud prior to being down-sampled, up-sampled, color converted, etc. Thus, a closed-loop color conversion module may receive a compressed version of a point cloud such as a decoder would receive and also a reference version of the point cloud before any distortion has been introduced into the point cloud due to sampling, compression, or color conversion.

[0272] In some embodiments, a closed-loop color conversion module, such as closed-loop color conversion module 410, may include a video decompression module, such as video decompression module 270, and a geometry reconstruction module, such as geometry reconstruction module 412. A video decompression module may decompress one or more video encoded image frames to result in decompressed image frames each comprising one or more patch images packed into the image frame. A geometry reconstruction module, such as geometry reconstruction module 412, may then generate a reconstructed point cloud geometry. A re-coloring module, such as re-coloring module 414, may then determine colors for points in the point cloud based on the determined reconstructed geometry. For example, in some embodiments, a nearest neighbor approach or other approach may be used to determine estimated color values for points of the point cloud based on sub-sampled color information, wherein a color value is not explicitly encoded for each point of the point cloud. Because there may be losses during the patching process, compression process, decompression process, and geometry reconstruction process, the geometry of the points in the reconstructed point cloud may not be identical to the geometry in the original point cloud. Due to this discrepancy, color compression techniques that rely on geometrical relationships between points to encode color values may result in colors that are slightly different when decoded and decompressed than the original colors. For example, if a color is to be determined based on color values of the nearest neighboring points, a change in geometry may cause a different nearest neighbor to be selected to determine the color value for the point at the decoder than was selected to encode a residual value at the encoder. Thus distortion may be added to the decoded decompressed point cloud.

[0273] If a color space conversion module does not account for this distortion that takes place when converting a point cloud into patches packed in an image frame and that takes place when encoding the image frames, the color space conversion module may select less than optimum color conversion parameters, such as luma and chroma values. For example, optimum color conversion parameters that cause a packed image frame in a first color space to closely match the packed image frame converted into a second color space may be different than optimum color conversion parameters when upstream and downstream distortions are accounted for.

[0274] In order to account for such distortions, a texture/attribute image color space conversion and re-sampling module, such as module 416, may take into account a difference between the “re-created” color values from re-coloring module 416 and the original color values from the original non-compressed reference point cloud when determining color conversion parameters for converting an image frame from a first color space, such as R’G’B’ 4:4:4 to YCbCr 4:2:0, for example. Thus, the color-converted and re-sampled texture/attribute images provided to video encoder 218 and 264, as shown in FIG. 4C may take into account distortion introduced at any stage of compression and decompression of a point cloud, and may utilize optimum color conversion parameters taking into account such distortion.

[0275] Such methods may result in considerably reduced distortion when reconstructing the point cloud representation, while maintaining the high compressibility characteristics of the 4:2:0 signal.

[0276] In some embodiments, conversion from 4:4:4 R’G’B’ to a 4:2:0 YCbCr representation is performed using a 3.times.3 matrix conversion of the form:

[ Y ’ Cb Cr ] = [ a YR a YG a YB a CbR a CbG a CbB a CrR a CrG a CrB ] [ R ’ G ’ B ’ ] ##EQU00007##

[0277] In the above matrix, Y’ is the luma component and Cb and Cr are the chroma components. The values of R’, G’, and B’ correspond to the red, green, and blue components respectively, after the application of a transfer function that is used to exploit the psycho-visual characteristics of the signal. The coefficients aYR through aCrB are selected according to the relationship of the red, green, and blue components to the CIE 1931 XYZ color space. Furthermore, the Cb and Cr components are also related to Y’ in the following manner:

Cb = B ’ – Y ’ alpha with alpha = 2 * ( 1 – a YB ) ##EQU00008## Cr = R ’ – Y ’ beta with beta = 2 * ( 1 – a YR ) ##EQU00008.2##

with also the following relationships:

a CbR = – a YR 2 * ( 1 – a YB ) ##EQU00009## a CbR = – a YG 2 * ( 1 – a YB ) ##EQU00009.2## a CbB = 0.5 ##EQU00009.3## a CrR = 0.5 ##EQU00009.4## a CrR = – a YG 2 * ( 1 – a YR ) ##EQU00009.5## a CrB = – a YB 2 * ( 1 – a YR ) ##EQU00009.6##

[0278] The process described above is followed by a 2.times. down-sampling horizontally and vertically of the chroma components, resulting in chroma components that are 4 times smaller, in terms of overall number of samples, 2.times. smaller horizontally and 2.times. smaller vertically, compared to those of luma. Such a process can help not only with compression but also with bandwidth and processing complexity of the YCbCr 4:2:0 signals.

[0279] In using such an approach quantization for the color components, as well as the down sampling and up sampling processes for the chroma components, may introduce distortion that could impact the quality of the reconstructed signals especially in the R’G’B’ but also in the XYZ (CIE 1931 domains). However, a closed loop conversion process, where the chroma and luma values are generated while taking into account such distortions, may considerably improve quality.

[0280] In a luma adjustment process, for example, the chroma components may be converted using the above formulations, additionally a down sampling and up sampling may be performed given certain reference filtering mechanisms. Afterwards, using the reconstructed chroma samples, an appropriate luma value may be computed that would result in minimal distortion for the luminance Y component in the CIE 1931 XYZ space. Such luma value may be derived through a search process instead of a direct computation method as provided above. Refinements and simplifications of this method may include interpolative techniques to derive the luma value.

[0281] Projected point cloud images can also benefit from similar strategies for 4:2:0 conversion. For example, closed loop color conversion, including luma adjustment methods may be utilized in this context. That is, instead of converting point cloud data by directly using the 3.times.3 matrix above and averaging all neighboring chroma values to generate the 4:2:0 chroma representation for the projected image, one may first project point cloud data/patches using the R’G’B’ representation on a 4:4:4 grid. For this new image one may then convert to the YCbCr 4:2:0 representation while using a closed loop optimization such as the luma adjustment method. Assuming that the transfer characteristics function is known, e.g. BT.709, ST 2084 (PQ), or some other transfer function as well as the color primaries of the signal, e.g. BT.709 or BT.2020, an estimate of the luminance component Y may be computed before the final conversion. Then the Cb and Cr components may be computed, down sampled and up sampled using more sophisticated filters. This may then be followed with the computation of the Y’ value that would result in a luminance value Yrecon that would be as close as possible to Y. If distortion in the RGB domain is of higher distortion, a Y’ value that minimizes the distortion for R’, G’, and B’ jointly, could be considered instead.

[0282] For point cloud data, since geometry may also be altered due to lossy compression, texture distortion may also be impacted. In particular, overall texture distortion may be computed by first determining for each point in the original and reconstructed point clouds their closest point in the reconstructed and original point clouds respectively. Then the RGB distortion may be computed for those matched points and accumulated across the entire point cloud image. This means that if the geometry was altered due to lossy compression, the texture distortion would also be impacted. Given that the texture may have been distorted, it may be desirable to consider geometry during closed loop conversion of chroma.

[0283] In some embodiments, the geometry is modified so that the relative sampling density in a given region of the point cloud is adjusted to be similar to other regions of the point cloud. Here the relative sampling density is defined as density of original points relative to the uniform 2D sampling grid.

[0284] Because the relative sampling density can vary within a given patch, this information can be used to guide the patch decomposition process as described above in regard to occupancy maps and auxiliary information, where patch approximation is used to determine local geometry. Furthermore, this information can be used to guide encoding parameters to achieve more uniform quality after compression. If a local region has higher relative sampling density, the encoder may code that region better through a variety of means. The variety of means may include: variable block size decision, Quantization Parameters (QPs), quantization rounding, de-blocking, shape adaptive offset (SAO) filtering, etc.

[0285] In some embodiments, the geometry information is first compressed according to a target bitrate or quality, and then it is reconstructed before generating the texture projected image. Then, given the reconstructed geometry, the closest point in the reconstructed point cloud is determined that corresponds to each point in the original point cloud. The process may be repeated for all points in the reconstructed point cloud by determining their matched points in the original point cloud. It is possible that some points in the reconstructed point cloud may match multiple points in the original point cloud, which would have implications in the distortion computation. This information may be used in the closed loop/luma adjustment method so as to ensure a minimized texture distortion for the entire point cloud. That is, the distortion impact to the entire point cloud of a sample Pr at position (x,y,z) in the reconstructed point cloud can be computed (assuming the use of MSE on YCbCr data for the computation):

D(Pr)=Doriginal(Pr)+Dreconstructed(Pr)

D(Pr)=Sum_matching(((Y_pr-Y_or(i)){circumflex over ( )}2+(Cb_pr-Cb_or(i)){circumflex over ( )}2+(Cr_pr-Cr_or(i)){circumflex over ( )}2)+sqrt((Y_pr-Y_or){circumflex over ( )}2+(Cb_pr-Cb_or){circumflex over ( )}2+(Cr_pr-Cr_or){circumflex over ( )}2)

[0286] In the above equation, Y_pr, Cb_pr, and Cr_pr are the luma and chroma information of point Pr, Y_or(i), Cb_or(i), and Cr_or(i) correspond to the luma and chroma information of all the points that were found to match the geometry location of point Pr from the original image, and Y_or, Cb_or, and Cr_or is the point that matches the location of point Pr in the original as seen from the reconstructed image.

[0287] If the distortion computation in the context of closed loop conversion/luma adjustment utilizes D(Pr), then better performance may be achieved since it not only optimizes projected distortion, but also point cloud distortion. Such distortion may not only consider luma and chroma values, but may instead or additionally consider other color domain components such as R, G, or B, luminance Y, CIE 1931 x and y, CIE 1976 u’ and v’, YCoCg, and the ICtCp color space amongst others.

[0288] If geometry is recompressed a different optimal distortion point may be possible. In that case, it might be appropriate to redo the conversion process once again.

[0289] In some embodiments, texture distortion, as measured as described below, can be minimized as follows: [0290] Let(Q(j)).sub.i.di-elect cons.{1, … , N} and (P.sub.rec(i)).sub.i.di-elect cons.{1, … , N.sub.rec.sub.} be the original and the reconstructed geometries, respectively. [0291] Let N and N.sub.rec be the number of points in the original and the reconstructed point clouds, respectively. [0292] For each point P.sub.rec(i) in the reconstructed point cloud, let Q*(i) be its nearest neighbor in the original point cloud and R(Q*(i)), G(Q*(i)), and B(Q*(i)) the RGB values associated with Q*(i). [0293] For each point P.sub.rec(i) in the reconstructed point cloud, let .sup.+(i)=(Q.sup.+(i,h)).sub.h.di-elect cons.{1, … , H(i)} be the set of point in the original point cloud that share P.sub.rec(i) as their nearest neighbor in the reconstructed point cloud. Note that .sup.+(i) could be empty or could have one or multiple elements. [0294] If .sup.+(i) is empty, then the RGB values R(Q*(i)), G(Q*(i)), and B(Q*(i)) are associated with the point P.sub.rec(i). [0295] If .sup.+(i) is not empty, then proceed as follows: [0296] Virtual RGB values, denoted R(.sup.+(i)), G(.sup.+(i)), and B(.sup.+(i)), are computed as follows:

[0296] R ( + ( i ) ) = 1 H ( i ) h = 1 H ( i ) R ( Q + ( i , h ) ) ##EQU00010## G ( + ( i ) ) = 1 H ( i ) h = 1 H ( i ) G ( Q + ( i , h ) ) ##EQU00010.2## B ( + ( i ) ) = 1 H ( i ) h = 1 H ( i ) B ( Q + ( i , h ) ) ##EQU00010.3## [0297] Note that R(.sup.+(i)), G(.sup.+(i)), and B(.sup.+(i)) correspond to the average RGB values of the points of .sup.+(i). [0298] The final RGB values R(P.sub.rec(i)), G(P.sub.rec(i)), and B(P.sub.rec(i)) are obtained by applying the following linear interpolation:

[0298] R(P.sub.rec(i))=wR(.sup.+(i))+(1-w)R(Q*(i))

G(P.sub.rec(i))=wR(.sup.+(i))+(1-w)G(Q*(i))

B(P.sub.rec(i))=wR(.sup.+(i))+(1-w)B(Q*(i)) [0299] The interpolation parameter w is chosen such that the following cost function C(i)* is minimized*

[0299] C ( i ) = max { 1 N h = 1 H ( i ) { ( R ( P rec ( i ) ) – R ( + ( i , h ) ) ) 2 + ( G ( P rec ( i ) ) – G ( + ( i , h ) ) ) 2 + ( B ( P rec ( i ) ) – B ( + ( i ) ) ) 2 } , 1 N rec { ( R ( P rec ( i ) ) – R ( Q * ( i ) ) ) 2 + ( G ( P rec ( i ) ) – G ( Q * ( i ) ) ) 2 + ( B ( P rec ( i ) ) – B ( Q * ( i ) ) ) 2 } } ##EQU00011## [0300] Note that by minimizing the cost C(i), the distortion measure as described below is minimized. [0301] Different search strategies may be used to find the parameter w [0302] Use the closed form solution described below. [0303] No search: use w=0.5. [0304] Full search: choose a discrete set of values (w.sub.i).sub.i=1 … W in the interval [0,1] and evaluate C(i) for these values in order to find the w*, which minimizes C(i). [0305] Gradient descent search: start with w=0.5. Evaluate E1(i), E2(i) and C(i). Store C(i) and w as the lowest cost and its associated interpolation parameter w. If E1(i)>E2(i), update w based on the gradient of E1(i), else use the gradient of E2(i). Re-evaluate E1(i), E2(i), and C(i) at the new value of w. Compare the new cost C(i) to the lowest cost found so far. If new cost is higher than the lowest cost stop, else update the lowest cost and the associated value of w, and continue the gradient descent, where R(P.sub.rec(i)), G(P.sub.rec(i)), and B(P.sub.rec(i)) are the three unknowns to be determined.

[0306] In some embodiments, the above process could be performed with other color spaces and not necessarily the RGB color space. For example, the CIE 1931 XYZ or xyY, CIE 1976 Yu’v’, YCbCr, IPT, ICtCp, La*b*, or some other color model could be used instead. Furthermore, different weighting of the distortion of each component could be considered. Weighting based on illumination could also be considered, e.g. weighting distortion in dark areas more than distortion in bright areas. Other types of distortion, that include neighborhood information, could also be considered. That is, visibility of errors in a more sparse area is likely to be higher than in a more dense region, depending on the intensity of the current and neighboring samples. Such information could be considered in how the optimization is performed.

[0307] Down sampling and up sampling of chroma information may also consider geometry information, if available. That is, instead of down sampling and up sampling chroma information without consideration to geometry, the shape and characteristics of the point cloud around the neighborhood of the projected sample may be considered, and appropriately consider or exclude neighboring samples during these processes. In particular, neighboring samples for down sampling or interpolating may be considered that have a normal that is as similar as possible to the normal of the current sample. Weighting during filtering according to the normal difference as well as distance to the point may also be considered. This may help improve the performance of the down sampling and up sampling processes.

[0308] It should be noted that for some systems, up sampling of the Cb/Cr information may have to go through existing architectures, e.g. an existing color format converter, and it might not be possible to perform such guided up sampling. In those cases, only considerations for down sampling may be possible.

[0309] In some embodiments, it may be possible to indicate in the bit stream syntax the preferred method for up sampling the chroma information. A decoder (included in an encoder), in such a case, may try a variety of up sampling filters or methods, find the best performing one and indicate that in the bit stream syntax. On the decoder side, the decoder may know which up sampling method would perform best for reconstructing the full resolution YCbCr and consequently RGB data. Such method could be mandatory, but could also be optional in some architectures.

[0310] Clipping as well as other considerations for color conversion, may also apply to point cloud data and may be considered to further improve the performance of the point cloud compression system. Such methods may also apply to other color representations and not necessarily YCbCr data, such as the YCoCg and ICtCp representation. For such representations different optimization may be required due to the nature of the color transform.

Example Objective Evaluation Method

[0311] A point cloud consists of a set of points represented by (x,y,z) and various attributes of which color components (y,u,v) are of importance. First, define the point v. It has as a mandatory position in a 3D space (x,y,z) and an optional color attribute c that has components r,g,b or y,u,v and optional other attributes possibly representing normal or texture mappings.

point v=(((x,y,z),[c],[a.sub.0 … a.sub.A]):x,y,z.di-elect cons.R,[c.di-elect cons.(r,g,b)|r,g,b.di-elect cons.N],[a.sub.i.di-elect cons.[0,1]]) (def. 1)

[0312] The point cloud is then a set of K points without a strict ordering:

Original Point Cloud V.sub.or={(v.sub.i):i=0 … K-1} (def. 2)

[0313] The point cloud comprises a set of (x,y,z) coordinates and attributes that can be attached to the points. The original point cloud Vor (420) will act as the reference for determining the quality of a second degraded point cloud Vdeg (424). Vdeg consists of N points, where N does not necessarily=K. Vdeg is a version of the point cloud with a lower quality possibly resulting from lossy encoding and decoding of Vor (e.g. operation 422). This can result in a different point count N.

Degraded Point Cloud V.sub.deg={(v.sub.i):i=0 … N-1} (def. 3)

[0314] The quality metric Q_(point cloud) is computed from Vor and Vdeg and used for assessment as shown in FIG. 4D for full reference quality metric 426.

[0315] Table 3, below, outlines the metrics used for the assessment of the quality of a point cloud, in some embodiments. The geometric distortion metrics are similar as ones used for meshes based on haussdorf (Linf) and root mean square (L2), instead of distance to surface. This approach takes the distance to the closest/most nearby point in the point cloud (see definitions 4, 5, 6, and 7) into account. Peak signal to noise ratio (PSNR) is defined as the peak signal of the geometry over the symmetric Root Mean Square (RMS/rms) distortion (def 8.). For colors, a similar metric is defined; the color of the original cloud is compared to the most nearby color in the degraded cloud and peak signal to noise ratio (PSNR) is computed per YUV/YCbCr component in the YUV color space (def. 10). An advantage of this metric is that it corresponds to peak signal to noise ratio (PSNR) in Video Coding. The quality metric is supported in the 3DG PCC software.

TABLE-US-00004 TABLE 3 Assessment criteria for assessment of the point cloud quality of Vdeg, Q.sub.point_cloud d_symmetric_rms Symmetric rms distance between the point clouds (def. 5.) d_symmetric_haussdorf Symmetric haussdorf distance between the clouds (def. 7.) psnr_geom Peak signal to noise ratio geometry (vertex positions) (def. 8.) psnr_y Peak signal to noise ratio geometry (colors Y) (def. 10) psnr_u Peak signal to noise ratio geometry (colors U) (as def. 10 rep. y for u) psnr_v Peak signal to noise ratio geometry (colors V) (as def. 10 rep. y for v) d rms ( V or , V deg ) = 1 K vo .di-elect cons. Vor vo – vd_nearest _neighbour 2 ##EQU00012## (def. 4) d.sub.symmetric_rms(V.sub.or, V.sub.deg) = max(d.sub.rms (V.sub.or, V.sub.deg), d.sub.rms(V.sub.deg, V.sub.or)) (def. 5) d.sub.haussdorf (V.sub.or, V.sub.deg) = max.sub.v.sub.o .sub..di-elect cons.V.sub.or, (||v.sub.o-v.sub.d_nearest_neighbour||.sub.2, v.sub.d is the point (def. 6) in Vdeg closest to v.sub.o (L2)) d.sub.symmetric_haussdorf(V.sub.or, V.sub.deg) = max(d.sub.haussdorf(V.sub.or, V.sub.deg), d.sub.haussdorf(V.sub.deg, V.sub.or) (def. 7) BBwidth = max((xmax-xmin), (ymax-ymin), (zmax-zmin) (def. 8) psnr.sub.geom = 10log.sub.10(|BBwidth||.sub.2.sup.2/(d.sub.symmetric rms(V)).sup.2) (def. 9) d y ( V or , V deg ) = 1 K vo .di-elect cons. Vor y ( vo ) – y ( v dnearest neighbour ) 2 ##EQU00013## (def. 10) psnr.sub.y = 10log.sub.10(|255||.sup.2/(d.sub.y(V.sub.or, V.sub.deg).sup.2) (def. 11)

[0316] In some embodiments, additional metrics that define the performance of a codec are outlined below in Table 4.

TABLE-US-00005 TABLE 4 Additional Performance Metrics Compressed size Complete compressed mesh size In point count K, the number of vertices in Vor Out point count N, number of vertices in Vdeg Bytes_geometry_layer Number of bytes for encoding the vertex positions Bytes_color_layer (opt) Number of bytes for encoding the colour attributes Bytes_att_layer (opt) Number of bytes for encoding the other attributes Encoder time (opt) Encoder time in ms on commodity hardware (optional) Decoder time (opt) Decoder time in ms on commodity hardware (optional)

Example Closed Form Solution

[0317] For each point P.sub.rec(i) in the reconstructed point cloud, let Q*(i) be its nearest neighbor in the original point cloud. For each point P.sub.rec (i) in the reconstructed point cloud, let (Q.sup.+(i,h)).sub.h.di-elect cons.{1, … , H(i)} be the set of point in the original point cloud that share P.sub.rec(i) as their nearest neighbor in the reconstructed point cloud. Let .sup.+(i) be the centroid of (Q.sup.+(i,h)).sub.h.di-elect cons.{1, … , H(i)}.

If H=0, then C(P.sub.rec(i))=C(Q*(i))

[0318] Denote as R-G-B vector C(P) associated with a given point P. In order to compute the color for a given P.sub.rec(i), we have the following formulation:

argmin C ( P rec ( i ) ) max { 1 N rec C ( P rec ( i ) ) – C ( Q * ( i ) ) 2 , 1 N h = 1 H C ( P rec ( i ) ) – C ( Q + ( i , h ) ) 2 } ##EQU00014## Where max { 1 N rec C ( P rec ( i ) ) – C ( Q * ( i ) ) 2 , h = 1 H C ( P rec ( i ) ) – C ( + ( i ) ) + C ( + ( i ) ) – C ( Q + ( i , h ) ) 2 } = max { 1 N rec C ( P rec ( i ) ) – C ( Q * ( i ) ) 2 , H N C ( P rec ( i ) ) – C ( + ( i ) ) 2 + 1 N h = 1 H C ( + ( i ) ) – C ( Q + ( i , h ) ) 2 + 2 N h = 1 H C ( P rec ( i ) ) – C ( + ( i ) ) , C ( + ( i ) ) – C ( Q + ( i , h ) ) } = max { 1 N rec C ( P rec ( i ) ) – C ( Q * ( i ) ) 2 , H N C ( P rec ( i ) ) – C ( + ( i ) ) 2 + 1 N h = 1 H C ( + ( i ) ) – C ( Q + ( i , h ) ) 2 } ##EQU00014.2##

[0319] Now denote D.sup.2=.SIGMA..sub.h=1.sup.H.parallel.C(.sup.+(i))-C(Q.sup.+(i,h)).paral- lel..sup.2,* so that*

argmin C ( P rec ( i ) ) max { 1 N rec C ( P rec ( i ) ) – C ( Q * ( i ) ) 2 , H N C ( P rec ( i ) ) – C ( + ( i ) ) 2 + D 2 N } . ##EQU00015##

[0320] Note: if H=1 then D.sup.2=0

[0321] Let C.sup.0(P.sub.rec(i)) be a solution of the previous minimization problem. It can be shown that C.sup.0(P.sub.rec(i)) could be expressed as:

C.sup.0(P.sub.rec(i))=wC(Q*(i))+(1-w)C(.sup.+(i))

[0322] Furthermore, C.sup.0(P.sub.rec(i)) verifies:

1 N rec wC ( Q * ( i ) ) + ( 1 – w ) C ( + ( i ) ) – C ( Q * ( i ) ) 2 = H N wC ( Q * ( i ) ) + ( 1 – w ) C ( + ( i ) ) – C ( + ( i ) ) 2 + D 2 N ( 1 – w ) 2 C ( + ( i ) ) – C ( Q * ( i ) ) 2 = w 2 HN rec N C ( Q * ( i ) ) – C ( + ( i ) 2 + D 2 N rec N ##EQU00016##

[0323] Let .delta..sup.2=.parallel.C(Q*(i)-C(.sup.+(i).parallel..sup.2** and**

r = N rec N ##EQU00017##

[0324] If .delta..sup.2=0, then C(P.sub.rec(i))=C(Q*(i))=C(.sup.+(i)

(1-w).sup.2.delta..sup.2=w.sup.2rH.delta..sup.2+rD.sup.2

.delta..sup.2+w.sup.2.delta..sup.2-2w.delta..sup.2=w.sup.2rH.delta..sup.- 2+rD.sup.2

.delta..sup.2(1-rH)w.sup.2-2.delta..sup.2w+(.delta..sup.2-rD.sup.2)=0

(rH-1)w.sup.2+2w+(.alpha.r-1)=0

[0325]* With*

.alpha. = D 2 .delta. 2 ##EQU00018##

[0326] if H=1, then w=1/2

[0327] if H>1

.DELTA.=4-4(rH-1)(.alpha.r-1)

.DELTA.=4-4(rH-1).alpha.r+4H-4

.DELTA.=4(H-(rH-1).alpha.r)

[0328] If .DELTA.=0

w = – 1 ( rH – 1 ) ##EQU00019##

[0329] If .DELTA.>0

w 1 = – 1 – ( H – ( Hr – 1 ) .alpha. r ) ( rH – 1 ) ##EQU00020## w 2 = – 1 + ( H – ( Hr – 1 ) .alpha. r ) ( rH – 1 ) ##EQU00020.2##

[0330] Where the cost C(i) is computed for both w1 and w2 and the value that leads to the minimum cost is retained as the final solution.

Compression/Decompression Using Multiple Resolutions

[0331] FIG. 5A illustrates components of an encoder that includes geometry, texture, and/or attribute downscaling, according to some embodiments. Any of the encoders described herein may further include a spatial down-scaler component 502, a texture down-scaler component 504, and/or an attribute down-scaler component 506 as shown for encoder 500 in FIG. 5A. For example, encoder 200 illustrated in FIG. 2A may further include downscaling components as described in FIG. 5A. In some embodiments, encoder 250 may further include downscaling components as described in FIG. 5A.

[0332] In some embodiments, an encoder that includes downscaling components, such as geometry down-scaler 502, texture down-scaler 504, and/or attribute down-scaler 506, may further include a geometry up-scaler, such as spatial up-scaler 508, and a smoothing filter, such as smoothing filter 510. In some embodiments, a reconstructed geometry image is generated from compressed patch images, compressed by video compression module 218. In some embodiments an encoder may further include a geometry reconstruction module (not shown) to generate the reconstructed geometry image. The reconstructed geometry image may be used to encode and/or improve encoding of an occupancy map that indicates patch locations for patches included in one or more frame images. Additionally, the reconstructed geometry image may be provided to a geometry up-scaler, such as spatial up-scaler 508. A geometry up-scaler may scale the reconstructed geometry image up to an original resolution or a higher resolution approximating the original resolution of the geometry image, wherein the original resolution is a resolution prior to downscaling being performed at spatial down-scaler 502. In some embodiments, the up-scaled reconstructed geometry image may be provided to a smoothing filter that generates a smoothed image of the reconstructed and up-scaled geometry image. The smoothing filter, such as smoothing filter 510, may approximate smoothing that would be applied via a smoothing filter of a decoder. This information may then be provided to the spatial image generation module 210, texture image generation module 212, and/or the attribute image generation module 214. These modules may adjust generation of spatial images, texture images, and/or other attribute images based on the reconstructed geometry images. For example, if a patch shape (e.g. geometry) is slightly distorted during the downscaling, encoding, decoding, and upscaling process, these changes may be taken into account when generating spatial images, texture images, and/or other attribute images to correct for the changes in patch shape (e.g. distortion). As an example, points of the point cloud represented in the reconstructed geometry image may be slightly moved as compared to the locations of the points in the original geometry image. In such circumstances, a texture image generation module, as an example, may take into account these distortions and adjust texture values assigned to the points of a corresponding texture image patch accordingly.

[0333] FIG. 5B illustrates components of a decoder 520 that includes geometry, texture, and/or other attribute upscaling, according to some embodiments. For example, decoder 520 includes texture up-scaler 512, attribute up-scaler 514, and spatial up-scaler 516. Any of the decoders described herein may further include a texture up-scaler component 512, an attribute up-scaler component 514, and/or a spatial image up-scaler component 516 as shown for decoder 520 in FIG. 5B.

[0334] FIG. 5C illustrates rescaling from the perspective of an encoder, according to some embodiments. In some embodiments, a point cloud may be scaled in both the point cloud domain (e.g. 3D domain prior to patch projection) and in a video level domain (e.g. by scaling 2D image frames comprising patch information). For example FIG. 5C illustrates a point cloud 522 of a person. An encoder, such as encoder 500, performs 3D scaling of the point cloud 522 in the point cloud domain to generate a downscaled point cloud 524. Patches generated based on downscaled point cloud 524 are packed into image frame 526. Additionally, downscaling is performed on the image frame 526 at the video level to reduce a resolution of the image frame. The additional downscaling results in a downscaled image frame 528 that is then encoded into a bit stream 530.

[0335] FIG. 5D illustrates rescaling from the perspective of a decoder, according to some embodiments. In some embodiments, a decoder, such as decoder 520, may receive a bit stream, such as bit stream 530. The decoder may decode the video encoded bit stream to generate one or more video image frames, such as image frame 532. The decoder may further upscale the image frame 532 to generate an up-scaled image frame 534. The decoder may then use a patch reconstruction method, as described above, to generate a reconstructed point cloud 536 from the patch information included in the up-scaled image frame 534. The decoder may also perform up-scaling in the 3D point cloud domain to scale up the reconstructed point cloud 536 to a similar size as the original point cloud. This process may result in an up-scaled reconstructed point cloud 538.

[0336] FIG. 5E illustrates an example open loop rescaling, according to some embodiments. In an open loop rescaling of an image frame, a geometry plane, and a texture or other attribute plane may be independently scaled, where geometry distortion is not taken into account when scaling the texture or other attribute information. For example, geometry image frame 540 may indicate depths of points of a point cloud relative to a projection plane and texture or attribute image frame 544 may represent respective attributes of the points of the point cloud projected on to the projection plane. As shown in FIG. 5E, in an open loop rescaling process, the geometry information and the attribute information may be independently scaled to generate down-scaled geometry image frame 542 and down-scaled texture or attribute image frame 546, respectively. Also, as shown in FIG. 5E, the downscaled geometry image frame 542 may be video encoded/compressed to generate a geometry bit stream and the downscaled attribute image frame 546 may be video encoded/compressed to generate a texture or attribute bit stream, such as a texture/attribute bit stream. For example, spatial down-scaler 502 may downscale the geometry image frame 540 and the texture down-scaler 504 may independently downscale the texture image frame 544. In some embodiments, attribute down-scaler 506 may downscale an attribute image frame independently of spatial down-scaler 502 and texture down-scaler 504. Because different down-scalers are used to downscale different types of image frames (e.g. spatial information, texture, other attributes, etc.), different downscaling parameters may be applied to the different types of image frames to downscale geometry different than texture or attributes.

[0337] FIG. 5F illustrates an example closed loop rescaling, according to some embodiments. In some embodiments, a closed loop rescaling process may be used by an encoder such as encoder 500 to determine distortion or other changes to geometry that may occur as part of a downscaling, encoding, decoding, and/or upscaling process. In some embodiments, such distortion may be accounted for when downscaling other attributes, such as texture. An encoder, such as encoder 500, receives a point cloud 548. The encoder generates a geometry image frame for the point cloud 548, for example an image frame comprising patches representing relative depths of the points, such as an original geometry image frame 550. A point cloud compression geometry mapper, which may include a decomposition into patches module 506, a packing module 208, and a spatial image generation module 210, etc., generates the original geometry image frame 550. A geometry down-scaler, such as spatial down-scaler 502 downscales the geometry image frame to generate downscaled geometry image frame 552. Note that “geometry plane” may be used to refer to geometry patch information, which may be included in an image frame only consisting of geometry patches as shown in FIG. 5F.

[0338] The downscaled geometry image frame 552 is compressed, for example by video compression module 218, and is converted into a geometry bit stream. In a closed loop process as shown in FIG. 5F, the geometry bit stream is decompressed at the encoder to generate a reconstructed geometry plane 554. The reconstructed geometry plane is then up-scaled, at the encoder, to generate an up-scaled reconstructed geometry plane 556.

[0339] The texture points of the original geometry image frame 550 are then mapped to the points of the reconstructed up-scaled geometry plane 556. Differences in locations of the points in the original geometry image frame and the re-constructed up-scaled geometry image frame are determined. Also, the points included in the geometry image frame 550 are adjusted to take into account distortion that may be introduced during the down-scaling, video compression, video-de-compression, and up-scaling processes. Additionally, this distortion may be taken into account by a point cloud compression (PCC) attribute/texture mapper to adjust texture values for points that are distorted during the down-scaling, video-compression, video-de-compression, and up-scaling process. Additionally, attribute values may also be adjusted to take into account geometry distortion. In this way, the texture and attribute points are mapped to the same points in the same locations as the decoder will encounter when reconstructing and up-scaling the geometry plane. Then, the encoder can take into account the distortion of the geometry plane that may occur due to downscaling, encoding, decoding, and upscaling.

[0340] The texture points mapped to the points of the reconstructed up-scaled geometry plane 556 may be used to generate an adjusted attribute/texture image frame 558. The adjusted attribute/texture image frame 558 may then be down-scaled to generate a down-scaled adjusted attribute/texture image frame 560. The down-scaled adjusted attribute/texture image frame 560 may then be video encoded and transmitted as an attribute/texture bit stream.

[0341] FIG. 5G illustrates an example closed loop rescaling with multiple attribute layers, according to some embodiments. In some embodiments, a similar process as described for FIG. 5F may be followed. However, multiple degrees of down-sampling may be performed for one or more attribute image frames being down-scaled. For example texture/attribute image plane 558 may not be downscaled at all (e.g. compression rate target 0), or may be downscaled according to a plurality of compression rate targets (e.g. compression rate targets 1-4) to generate down-scaled versions of the adjusted attribute/texture image frame 562. In such embodiments, a compression rate target may be dynamically adjusted, for example based on network conditions, processing capacity, etc.

[0342] FIG. 5H illustrates an example of video level spatio-temporal scaling, according to some embodiments. In some embodiments, a similar process as described in FIGS. 5C and 5D may be performed using video level spatio-temporal downscaling and upscaling. For example, a frame rate (e.g. a number of frames generated per unit time) may be adjusted up or down in order to improve compression efficiency. In such embodiments spatial temporal adjustments may be made instead of resolution scaling and/or in addition to resolution scaling. For example, point clouds 564 at sequential movements in time may be down-scaled in resolution and/or frame rate in the 3D domain to generate down-scaled point clouds 566. The point clouds 566 may be projected onto a patch plane and image frames 568 may be generated. Note image frames 568 may be geometry image frames or attribute image frames, or both. Additionally, video level spatio-temporal down-scaling may be applied to reduce a resolution of the image frames 568 and/or reduce a frame-rate of the image frames 568 to generate down-scaled image frames 570. Note FIG. Note FIG. 5H illustrates both frame-rate down-scaling (e.g. spatial temporal down-scaling) and resolution down-scaling. However, in some embodiments, spatio-temporal down-scaling may be performed without performing resolution down-scaling. The spatio-temporal and/or resolution down-scaled image frames 570 may then be video-encoded to generate bit stream 572.

[0343] FIG. 5H also illustrates an encoder receiving bit-stream 572. A video-decoding component of the decoder may video-decode the bit stream 572 to generate down-scaled image frames 574. The decoder may also perform video level-spatio temporal upscaling to interpolate between the down-scaled image frames to generate up-scaled image frame 576. For example, down-scaled image frames, as an example includes two image frames per unit of time, whereas up-scaled image frames 576 includes a third image frame that has been generated in the 2D video domain by interpolating, and/or using other video-spatial intra frame compression techniques, such as motion vectors, etc. to generate the third image frame. The up-scaled image frames 576 may then be used to generate reconstructed point clouds 578. Note that three reconstructed point clouds have been generated based on up-scaled image frames 576. Optionally, the decoder may further upscale the reconstructed point clouds 578 either temporally or size-wise, or both, to generate up-scaled point clouds 580.

[0344] FIG. 5I illustrates an example closed loop rescaling with spatiotemporal scaling, according to some embodiments. For example, point clouds 582 may be used to generate original geometry image frames 584. The original geometry frames may further be down-scaled temporally and/or size-wise to generate down-scaled geometry image frames 586, which may have a different frame rate, such as fg frames per second, as compared to fo frames per second for original geometry image frames 584, where fg is less than fo. The down-scaled geometry image frames 586 may also have a smaller size than the original geometry image frames 584, such as height and width “g” as compared to height and width “o” of original geometry image frames 584. The downs-scaled geometry image frames may further be video-encoded/compressed to generate geometry bit stream 596. In a closed loop compression procedure, the geometry bit-stream 596 may further be video-decompressed/decoded at the encoder to generate reconstructed down-scaled geometry images 588, which may have a similar frame rate and size as down-scaled geometry images 586. The encoder may further apply a similar spatio-temporal and/or size-based up-scaling algorithm as would be executed at a decoder to generate up-scaled reconstructed image frames 590, which may have a similar frame rate and size as original geometry image frames 584. In some embodiments, the encoder may further adjust the original geometry image frames and repeat the process to reduce distortion. In some embodiments, geometry bit-stream 596 communicated out of the encoder may be based on adjusted geometry image frames 584.

[0345] In some embodiments, a point cloud compression (PCC) texture/attribute mapper may further adjust attribute/texture values of attribute/texture image frames based on distortion introduced due to the down-scaling, video-encoding, video-decoding, and up-scaling of the geometry image frames. For example, adjusted attribute/texture image frames 592 may be generated. The adjusted attribute/texture image frames 592 may further be down-scaled at generate down-scaled adjusted attribute/texture image frames 594, which may in turn be video-encoded to generate texture/attribute bit stream 598. While not illustrated for spatio-temporal down-scaling, a similar process as described in FIG. 5G may be performed, wherein a level of spatio-temporal down-scaling to be applied is determined based on an available bit rate to communicate the compressed point cloud.

[0346] As discussed above, methods of compressing point cloud video data may use conventional video codecs as well as auxiliary information that can help describe and reconstruct the point cloud information. The encoder and decoder diagrams of how that process is performed is shown in at least FIGS. 5A and 5B, respectively. Also, FIGS. 6A-61 further illustrate how relationships between geometry image frames and attribute/texture image frames may be used to improve compression efficiency and reconstruction accuracy. Similar techniques may be combined with up-scaling and down-scaling techniques as described herein in FIGS. 5A-5L.

[0347] As discussed above, the process of generating the image frames segments the point cloud into multiple 2D projected images/videos, each representing different types of information. This process is performed by segmenting the point cloud first into multiple patches that permit efficient projection of the entire 3D space data onto 2D planes. Each patch is associated with information such as geometry (also referred to herein as “spatial information”), texture, and other attributes if they are available. Such information is then copied at the co-located locations of the image frames on separate image sequences with each now containing only the geometry information, the texture information, and any other remaining attributes respectively. Auxiliary information that contains the patch information as well as an occupancy map that dictates which areas in these projected image sequences correspond to actual point cloud data and which are unoccupied, e.g. may contain no data or dummy data, are also provided. Compression is then applied on such information using different strategies. Auxiliary information, for example, is entropy coded, while occupancy maps may be down-converted and encoded using either conventional codecs or other methods such as run length compression. The separate projected image sequences on the other hand are compressed using conventional codecs. This results in a collection of multiple sub streams, e.g. a geometry sub stream, texture and attribute sub streams, as well as occupancy and auxiliary information sub streams.

[0348] As described above, all sub streams except the occupancy map are expected to be of the same resolution. Each point in the geometry sub stream essentially corresponds to a point in the final 3D reconstructed point cloud. In some embodiments, it is permitted for the signal to be encoded at a different resolution than the original representation. Also, in some embodiments, offsetting as well as rotating the point cloud is also possible. Seeing things from the encoder perspective, this is done by signaling in the stream header additional metadata that would identify the scaling, offset, and rotation that should be applied onto the original point cloud data prior to projecting it onto the target video planes. From the decoder perspective, these parameters are used after the reconstruction of a first 3D point cloud representation and utilized to generate the final 3D point cloud representation. In such a scheme, both geometry and attribute/texture video data are signaled at the same resolution as specified in the point cloud header. Per patch metadata including scaling factors and rotation parameters are also supported in such a scheme, with scaling though now applied on each projected patch independently.

[0349] However, in some embodiments, this scheme may be further extended by providing additional resolution scaling flexibility in the encoded streams. In particular, not only may the scaling be applied in 3D space or per patch, but in some embodiments scheme scaling may be applied on the entire projected point cloud video data. This may permit use of “conventional” 2D