VVC intra-frame prediction fast mode decision-making method suitable for V-PCC

Abstract

The invention relates to the technical field of three-dimensional point cloud compression and video coding, in particular to a VVC intra-frame prediction fast mode decision-making method suitable for V-PCC, and the method comprises the steps: constructing model input; extracting content features and boundary features input by the model, and splicing to obtain shared features; performing intra-frame mode type prediction on the shared features, performing secondary prediction according to a prediction type judgment result, and outputting candidate modes; the encoder performs final encoding decision according to the candidate mode; according to the intra-frame mode searching method and device, the intra-frame angle prediction, the matrix intra-frame prediction and the enabling prediction of the intra-frame sub-partition tool are jointly decided in the intra-frame prediction stage, and the candidate mode which is more likely to appear is pushed to the encoder according to the prediction probability, so that the intra-frame mode searching complexity is reduced, the coding efficiency is improved, and the coding time is shortened.

Description

A Fast Mode Decision Method for VVC Intra-Frame Prediction Applicable to V-PCC Technical Field

[0001] This application relates to the field of 3D point cloud compression and video coding technology, specifically to a fast mode decision method for VVC intra-frame prediction applicable to V-PCC. Background Technology

[0002] The next-generation video coding standard Versatile Video Coding (VVC / H.266) has extremely high compression efficiency. During intra-frame coding, VVC introduces richer intra-frame prediction mechanisms and tools, such as extending to 67 angle modes and introducing tools such as Multiple Reference Line (MRL), Intra-Frame Sub-Partition (ISP), and Matrix Intra-Frame Prediction (MIP). This makes intra-frame prediction mode comparison and rate-distortion optimization (RDO) search one of the main sources of coding time consumption. The flowchart of the VPCC intra-frame coding mode based on VVC in the existing technology is shown in Figure 1.

[0003] To reduce complexity, existing fast intra-frame mode decision-making techniques for VVC natural video / image coding have proposed approaches such as reducing the candidate mode list (CML), accelerating the coarse selection (RMD), and reducing the RDO search, including pruning methods based on statistical features or machine learning. However, these VVC fast intra-frame mode decision-making techniques mainly focus on establishing datasets and statistical patterns for natural video content, whose feature distribution differs significantly from that of point cloud projection videos. The geometric, attribute, and occupancy maps generated by point cloud projection often contain a large number of blank / invalid regions introduced by processes such as filling, and the content is sparse, highly structured, and has different texture statistical characteristics than natural images. This makes it difficult to directly transfer existing fast intra-frame mode decision-making methods for natural videos to V-PCC / TMC2 scenarios. Specifically, existing technologies have the following shortcomings:

[0004] 1. Existing V-PCC fast coding research focuses more on optimizing other coding stages, and lacks support for joint fast decision-making of angle modes and tools such as ISP / MIP / MRL in VVC intra-frame prediction; ISP (intra-frame sub-partition) increases the computational load by 20-40%, but only 5-15% of CUs actually need it; MIP (matrix prediction) increases the computational load by 15-30%, but its usage rate is only 3-10%; Figures 3, 4, and 5 show the intra-frame mode distribution of CUs with different occupancy types in the near layer of attribute video, showing the angle mode distribution and the activation ratio of MIP and ISP. It can be seen that the proportion of MIP and ISP is very small, but the encoder has to make a comparison judgment to determine whether to use them each time;

[0005] 2. The sparsity of point cloud projection content and the differences in multiple view types are not fully modeled: Geometric graphs, attribute graphs and occupancy graphs differ significantly in texture distribution, edge structure and effective pixel density. Directly using natural video statistics or features can lead to mispruning or conservative pruning, affecting coding performance or complexity gains.

[0006] 3. Insufficient engineering usability of multi-tool joint decision-making: VVC intra-frame prediction involves multiple tools such as angle mode, MRL, ISP, and MIP. Without a unified probability evaluation and candidate list pruning strategy, it is difficult to achieve fast decision-making that balances real-time performance and rate-distortion performance in the actual coding process of TMC2.

[0007] Therefore, in order to address the problem that the intra-frame mode search complexity is too high when V-PCC / TMC2 point cloud projection video is encoded using VVC, and that existing fast intra-frame decision-making methods based on natural video guidance are difficult to directly adapt, it is urgent to propose a fast intra-frame mode decision-making method that combines the characteristics of point cloud projection video, can simultaneously handle angle modes and tools such as ISP and MIP, and reduces complexity while ensuring compression efficiency. Summary of the Invention

[0008] In view of this, this application discloses a fast mode decision method for VVC intra-frame prediction applicable to V-PCC, to solve the problems in the prior art, including:

[0009] S1, Input for constructing the model;

[0010] S2. Extract the content features and boundary features of the model input, and concatenate them to obtain shared features;

[0011] S3. Perform intra-frame mode type prediction on the shared features, and perform secondary prediction based on the prediction type discrimination result to output candidate modes;

[0012] S4. The encoder makes the final encoding decision based on the candidate patterns.

[0013] The beneficial effects of this application include:

[0014] In this application, when making intra-frame mode decisions for geometric video blocks in a point cloud projection video encoder, the corresponding occupied block features are introduced as auxiliary inputs. By combining the characteristics of point cloud projection video, which is highly sparsity, has obvious structural features, and has large differences between different projection view types, feature information adapted to point cloud characteristics is introduced to improve the perception of effective point distribution and structural boundaries. This makes the intra-frame mode decision more consistent with the characteristics of point cloud projection content and avoids the problem of insufficient adaptability caused by directly applying natural video methods.

[0015] When making intra-frame mode decisions for attribute video blocks, corresponding geometric block features and occupied block features are introduced as auxiliary inputs, so that attribute prediction can make full use of the spatial structure information and effective point distribution information of the point cloud, thereby improving the consistency and accuracy of intra-frame mode decisions.

[0016] By enabling joint decision intra-angle prediction, matrix intra-frame prediction (MIP), and intra-fractional sub-partitioning (ISP) tools, the complexity of intra-frame mode search is reduced. During the training phase, the encoder rate-distortion optimization (RDO) results are used as supervision signals to guide candidate mode selection during the inference phase. This aligns the model training objective with the encoder's optimal decision on rate-distortion cost, making the model output approximate the RDO search results. Only candidate selection and complexity control are performed, and the final encoding decision is still determined by the RDO results, thus achieving fast decision-making for VVC intra-frame mode prediction.

[0017] This application improves the practicality and real-time performance of point cloud video coding by jointly modeling and hierarchically predicting multiple modes and tools in intra-frame prediction, while ensuring that coding efficiency is not significantly reduced, thereby reducing the number of redundant mode searches and lowering coding complexity. This provides a new design approach for those skilled in the art. Attached Figure Description

[0018] Figure 1 is a schematic diagram of the VPCC intra-frame coding mode based on VVC in the prior art;

[0019] Figure 2 is a schematic diagram of the workflow of the multi-task prediction neural network model in the embodiments of this application;

[0020] Figure 3 is a schematic diagram of attribute video not occupying CU in the near layer in the embodiment of this application;

[0021] Figure 4 is a schematic diagram of the near-layer portion of the attribute video occupying the CU in an embodiment of this application;

[0022] Figure 5 is a schematic diagram showing that the attribute video in the near layer occupies all the CUs in the embodiment of this application;

[0023] Figure 6 is a flowchart illustrating the VVC intra-frame prediction fast mode decision method applicable to V-PCC in the embodiments of this application;

[0024] Figure 7 is a schematic diagram of the backbone network structure of the multi-task prediction neural network model in the embodiment of this application;

[0025] Figure 8 shows the coding performance test results in the embodiments of this application. Detailed Implementation

[0026] To make the objectives, technical solutions, features, and advantages of this application clearer and to enable those skilled in the art to better understand the technical solutions of this application, the following detailed description of this application is provided in conjunction with the accompanying drawings and embodiments.

[0027] Example 1:

[0028] This embodiment includes a fast mode decision method for VVC intra-frame prediction applicable to V-PCC, implemented based on a multi-task prediction neural network model, including:

[0029] S1. Construct model input; Model input includes: coding unit extension block, intra-frame prediction reference pixel sequence and its effective length, size information of the current coding unit and its size group identifier, and quantization parameter QP of the current coding unit.

[0030] Furthermore, the extended coding unit block includes: the current coding unit pixel and its adjacent reference pixels above, to the left and to the top left corner; the intra-frame prediction reference pixel sequence originates from the same intra-frame prediction reference region as the extended block, and its original length is related to the coding unit size; in order to achieve unified modeling between coding units of different sizes, the one-dimensional reference pixel sequence is mapped to a fixed-length sequence of a preset length (e.g., 128) through interpolation resampling, which serves as the input for boundary feature extraction.

[0031] The size group identifier is obtained by pre-grouping the coding units (CUs) based on differences in size and shape, including: square coding unit group, horizontal rectangular coding unit group, and vertical rectangular coding unit group; the square coding unit group includes: square coding unit group 1 (w=8), square coding unit group 2 (w=4), and square coding unit group 3 (w>8); a corresponding size identifier (size_id) is assigned to each CU, and the size identifier is mapped to a size feature vector for subsequent conditional feature modulation. Sub-modules of the multi-task prediction neural network model are constructed for different size groups.

[0032] S2. Extract the content features and boundary features of the model input, and concatenate them to obtain shared features, including:

[0033] S21. Extract content features from the extended blocks of the coding unit using a two-dimensional convolutional network;

[0034] S22. Extract the boundary features of the intra-frame prediction reference pixel sequence through a one-dimensional convolutional network; resample the reference pixel sequence used for intra-frame prediction, convert it into a fixed-length one-dimensional sequence, and encode it using a one-dimensional convolutional network to obtain the boundary feature vector.

[0035] S23. Through the FiLM (Feature-wise Linear Modulation) mechanism, modulation parameters are generated based on the size group identifier and quantization parameter QP of the current coding unit. The modulation parameters are applied to the residual block of the two-dimensional convolutional network so that the content features can adapt to different coding sizes and coding conditions.

[0036] Specifically, the FiLM mechanism generates modulation parameters γ and β based on the coding unit size characteristics and quantization parameters QP, and applies these parameters to the residual block to adaptively scale and shift the feature map, thereby improving the model's adaptability under different sizes and coding conditions. By fusing learning-guided probabilities and statistical-guided probabilities to evaluate and prune candidate modes, the number of candidate modes searched for entry rate distortion optimization is reduced while ensuring that coding efficiency is not significantly reduced, achieving a reasonable balance between complexity and coding performance.

[0037] S24. Utilize boundary features to generate gating vectors and perform weighted modulation on content features to enhance the correlation between intra-frame prediction and boundary reference. This step enables boundary features to guide the selection and enhancement of content features, thereby improving the modeling capability of intra-frame prediction and boundary consistency.

[0038] S25. Pool and concatenate the content features and boundary features to obtain shared features for subsequent prediction.

[0039] S3. Perform intra-frame mode type prediction on the shared features, and perform secondary prediction based on the prediction type discrimination result to output candidate modes.

[0040] The intra-frame mode type prediction outputs the prediction type discrimination result between angle prediction and matrix intra-frame prediction (MIP), indicating whether the current coding unit uses the angle prediction tool or the MIP tool. When determining whether to enable matrix intra-frame prediction mode, the SATD cost value calculated based on the intra-frame angle prediction mode DC and mode PLANA is introduced as an auxiliary feature: the SATD cost value includes SATD_DC and SATD_PLANAR, corresponding to the intra-frame angle prediction mode DC and intra-frame angle prediction mode PLANA, respectively. SATD_DC and SATD_PLANAR are used as fast distortion metrics to characterize the prediction effectiveness of traditional angle prediction on the current coding unit, thereby assisting in determining whether to enable MIP mode. This avoids mistakenly enabling MIP when traditional angle prediction already has good results, or excessively filtering angle modes when selecting MIP.

[0041] The second-level prediction is performed based on the prediction type discrimination result. If the discrimination result is an angle mode, the distribution probability of all intra-frame angle prediction modes and the activation probability of the intra-frame sub-partition (ISP) tool are output. If the discrimination result is a MIP mode, the probability distribution of the MIP mode under the corresponding size is output. When the coding unit size is 4×4 or 8×8, 16 MIP modes are output, otherwise 35 MIP modes are output, thereby reducing computational overhead.

[0042] Furthermore, the patterns are sorted according to the probability distribution, and candidate patterns are output.

[0043] Candidate modes are used to guide encoder selection, rather than directly replacing the encoder's rate-distortion optimization calculation. Only a few candidate modes with high prediction probabilities are selected from the corresponding branches, and then standard rate-distortion optimization search is performed on these candidate modes, thereby reducing intra-frame prediction complexity while ensuring standard compatibility.

[0044] S4. The encoder makes the final encoding decision based on the candidate patterns.

[0045] The method designed in this application is mainly used for candidate screening and complexity control, while the final coding decision is still determined by the RDO results.

[0046] Furthermore, the method designed in this application is based on a pre-trained multi-task prediction neural network model. During the training phase, an offline rate-distortion optimization search is performed on each encoding unit. The corresponding rate-distortion cost RDCost is calculated for different combinations of prediction tools, modes, and ISP configurations, and the configuration with the minimum RDCost is selected as the optimal decision. Based on the optimal decision, prediction type labels, angle mode labels, MIP mode labels, and ISP labels are generated to ensure that the model training objective is consistent with the optimal rate-distortion strategy actually adopted by the encoder.

[0047] Furthermore, training employs a hierarchical multi-task joint loss, with the total loss function formula as follows:

[0048]

[0049]

[0050] in, The loss represents the model prediction loss, including the first-level loss term and the second-level loss term. This indicates the predicted loss for ISP (Intra Sub-Partition). and These represent the weighting coefficients. In this embodiment, the following are selected: =1.0、 =0.5, and in some embodiments, the complexity and accuracy requirements can be adjusted according to the encoder side; This indicates the first level of loss. This represents the angle pattern classification loss. The weights represent the classification loss of the MIP pattern. This represents the MIP pattern classification loss. and Together they constitute the second level of loss item.

[0051] The first-level loss term is used to distinguish whether a sample belongs to the angle pattern or the MIP pattern, constraining the model to correctly distinguish between angle predictions and MIP predictions. It employs cross-entropy loss.

[0052]

[0053] Where B represents the batch size; For the sample The prediction type is logits, angular mode = 0, and MIP mode = 1; ∈{0,1} represents the true type label; Represents the cross-entropy loss function

[0054] Furthermore, when class imbalance exists, the first-level loss can employ a variant of Focal Loss, with the following formula:

[0055]

[0056] in , Represents the balance factor. This represents the focusing parameter; in this embodiment, the following parameters are selected: = 0.25, =2.

[0057] The second-level loss term is the angle pattern classification loss. MIP pattern classification loss constitute.

[0058] The angle pattern classification loss constrains the prediction accuracy of the angle pattern distribution in the angle prediction scenario. It is calculated only for samples corresponding to the "angle prediction of the optimal RDCost decision," avoiding imposing invalid constraints on irrelevant prediction branches. For the angle sample set... The angle pattern classification loss is:

[0059]

[0060]

[0061] in, The angular pattern is represented by logits (67 angular patterns). Indicates the true angle mode label. Indicates sample weights, For the sample The cost of rate distortion; , Minimum / maximum rate distortion cost within the batch; It is a numerically stable term; and In this embodiment, the weight range is selected. =0.5, =1.5. The angle branch loss is weighted according to the sample rate-distortion cost, making the training target closer to the rate-distortion optimal decision.

[0062] The MIP pattern classification loss is used to constrain the prediction accuracy of the MIP pattern distribution in the MIP prediction scenario. It is calculated only for samples corresponding to the "MIP prediction of the optimal RDCost decision" to avoid invalid constraints. For the MIP sample set... The MIP pattern classification loss is:

[0063]

[0064] in, This indicates MIP mode logits. =16 or 35, which can vary depending on the CU size. This indicates a true MIP mode tag.

[0065] The ISP pattern classification loss, used to constrain the prediction results for whether the ISP tool is enabled or not, is formulated as follows:

[0066]

[0067] in, This indicates that the ISP predicts the logit. The ISP tag indicates that 0 means no ISP is used and 1 means an ISP is used. This represents the binary cross-entropy loss function.

[0068] Furthermore, the total loss function is obtained by combining the various losses, and the formula is as follows:

[0069]

[0070] Using the optimal decision of RDCost as a supervision signal and training with the aforementioned hierarchical multi-task joint loss, the model output can approximate the rate-distortion optimization result of the encoder, while avoiding imposing invalid learning constraints on irrelevant branches, thereby improving the accuracy of intra-frame mode selection and coding efficiency.

[0071] Furthermore, Figure 2 shows a schematic diagram of the workflow of the multi-task prediction neural network model in this embodiment. In the figure, P1 and P2 are the occurrence frequency and the model output probability. Probability estimation refers to pruning by combining the distribution probability and the model output probability. At the same time, the statistical probabilities of geometric video and projection video are different.

[0072] This application introduces corresponding occupancy block features as auxiliary input when making intra-frame mode decisions for geometric video blocks in a point cloud projection video encoder, thereby improving the perception of effective point distribution and structural boundaries. When making intra-frame mode decisions for attribute video blocks, it simultaneously introduces corresponding geometric block features and occupancy block features as auxiliary input, enabling attribute prediction to fully utilize the spatial structure information and effective point distribution information of the point cloud, thereby improving the consistency and accuracy of intra-frame mode decisions. By enabling joint decision-making intra-frame angle prediction, matrix intra-frame prediction (MIP), and intra-frame sub-partitioning (ISP) tools, the complexity of intra-frame mode search is reduced. During the training phase, the encoder rate-distortion optimization (RDO) results are used as a supervision signal to guide candidate mode selection during the inference phase, thereby achieving fast decision-making for VVC intra-frame mode prediction.

[0073] Example 2:

[0074] This embodiment includes a fast mode decision method for VVC intra-frame prediction applicable to V-PCC, as shown in Figure 6. The difference from Embodiment 1 is that the model input further includes auxiliary features of geometric blocks and occupancy blocks. When intra-coding attribute video, the auxiliary features of geometric blocks and occupancy blocks are used to participate in intra-frame mode decision-making, and the geometric blocks and occupancy blocks guide the attribute block encoding. Geometric video only requires occupancy block guidance; the geometric blocks and occupancy blocks jointly guide the attribute blocks.

[0075] Specifically, occupancy blocks reflect whether the 3D pixel occupancy of geometric attributes is a useful pixel. Occupancy blocks are used to guide geometric video, while geometric variance is directly related to attribute information. Geometric blocks and occupancy blocks jointly guide attribute block encoding. When predicting a video coding unit for a specific attribute, the occupancy video block and geometric video block corresponding to that attribute coding unit in spatial location are first obtained, and occupancy features and geometric features are extracted respectively. Occupancy features are used to characterize the existence of valid points in the current region, and geometric features are used to characterize the spatial structure information of the point cloud. Auxiliary features will help the encoder make fast intra-frame mode decisions.

[0076] Figure 7 shows a schematic diagram of the backbone network structure of the multi-task prediction neural network model in this embodiment. FiLM is the name of the multi-task prediction neural network model in this embodiment. Figure 8 shows the coding time change and coding loss after the point cloud fast coding unit partitioning decision in this embodiment. The test condition is full-frame configuration. The coding loss is measured by the Bjontegaard Delta-Rate (BDBR), a commonly used video coding metric, and obtained using the table calculation tool provided by TMC2. Figure 8 shows the average value after testing the loot, basketball, dancer, and soldier sequences. In the figure, ∆T represents the percentage of coding time saved, and the formula is:

[0077]

[0078] Where 'i' represents the video type, including geometric video and attribute video. This represents the time consumed by the encoder using the design method of this application when the video type is i; This indicates the time consumed by the original encoder TMC2-v18.0 when the video type is i.

[0079] It can be seen that the method designed in this application can effectively reduce coding complexity, shorten coding time, and control coding loss within a negligible range.

[0080] Finally, it should be noted that the above description only depicts some embodiments of this application. For those skilled in the art, various changes, modifications, substitutions, and variations can be conceived of these embodiments without departing from the principles and spirit of this application. The scope of protection of this application is defined by the appended claims and their equivalents, and all the above-mentioned behaviors should be covered within the scope of protection of this application.

Claims

1. A fast mode decision-making method for VVC intra-frame prediction suitable for V-PCC, implemented based on a multi-task prediction neural network model, characterized in that, include: S1. Construct model input; model input includes: coding unit extension block, intra-frame prediction reference pixel sequence and its effective length, size information of the current coding unit and its size group identifier, and quantization parameters of the current coding unit; S2. Extract content features and boundary features from the model input and concatenate them to obtain shared features; S3. Perform intra-frame mode type prediction on the shared features, and perform secondary prediction based on the prediction type discrimination result to output candidate modes; the intra-frame mode type prediction outputs the prediction type discrimination result between angle prediction and matrix intra-frame prediction; the secondary prediction based on the prediction type discrimination result, if the discrimination result is angle mode, outputs the distribution probability of all intra-frame angle prediction modes and the activation probability of the intra-frame sub-partition tool; if the discrimination result is matrix intra-frame prediction mode, outputs the probability distribution of matrix intra-frame prediction mode under the corresponding size; sort the modes according to the probability distribution and output candidate modes; S4. The encoder makes the final encoding decision based on the candidate modes.

2. The fast mode decision method for VVC intra-frame prediction applicable to V-PCC according to claim 1, characterized in that, The size group identifier is obtained by pre-grouping the coding units (CU) according to differences in size and shape, including: square coding unit group, horizontal rectangular coding unit group, and vertical rectangular coding unit group; and assigning a corresponding size identifier to each CU.

3. The fast mode decision method for VVC intra-frame prediction applicable to V-PCC according to claim 1, characterized in that, The extraction of content features and boundary features from the model input includes: S21, extracting content features from the extended block of the coding unit through a two-dimensional convolutional network; S22, extracting boundary features from the intra-frame prediction reference pixel sequence through a one-dimensional convolutional network; S23, generating modulation parameters based on the size group identifier and quantization parameters of the current coding unit through the FiLM mechanism, and applying the modulation parameters to the residual block of the two-dimensional convolutional network; S24, generating a gating vector using the boundary features and performing weighted modulation on the content features; S25, pooling and concatenating the content features and boundary features to obtain shared features for subsequent prediction.

4. The fast mode decision method for VVC intra-frame prediction applicable to V-PCC according to claim 1, characterized in that, When determining whether to enable matrix intra-frame prediction mode, the SATD value calculated based on intra-frame angle prediction mode DC and mode PLANA is used as an auxiliary feature. The SATD value includes SATD_DC and SATD_PLANAR, which correspond to intra-frame angle prediction mode DC and intra-frame angle prediction mode PLANA, respectively.

5. The fast mode decision method for VVC intra-frame prediction applicable to V-PCC according to claim 1, characterized in that, The probability distribution of the matrix intra-frame prediction mode corresponding to the output size is as follows: when the coding unit size is 4×4 or 8×8, 16 MIP modes are output; otherwise, 35 MIP modes are output.

6. The fast mode decision method for VVC intra-frame prediction applicable to V-PCC according to claim 1, characterized in that, The multi-task prediction neural network model is pre-trained before use. During the training phase, an offline rate-distortion optimization search is performed for each coding unit. The corresponding rate-distortion cost RDCost is calculated for different combinations of prediction tools, modes, and ISP configurations, and the configuration with the minimum RDCost is selected as the optimal decision. Based on the optimal decision, prediction type labels, angle mode labels, MIP mode labels, and ISP labels are generated.

7. The fast mode decision method for VVC intra-frame prediction applicable to V-PCC according to claim 6, characterized in that, Training employs a hierarchical multi-task joint loss mechanism, with the total loss function formula as follows: ;in, Indicates the model prediction loss. This represents the ISP's predicted loss, used to constrain the prediction results depending on whether the ISP tool is enabled or not. and These represent the weighting coefficients.

8. The fast mode decision method for VVC intra-frame prediction applicable to V-PCC according to claim 7, characterized in that, The pattern prediction loss includes a first-level loss term and a second-level loss term; the first-level loss term is used to distinguish whether a sample belongs to the angular pattern or the MIP pattern; the second-level loss term consists of the angular pattern classification loss and the MIP pattern classification loss; the formula for the pattern prediction loss is: ;in, This indicates the first level of loss. This represents the angle pattern classification loss. The weights represent the classification loss of the MIP pattern. This represents the MIP pattern classification loss.

9. The fast mode decision method for VVC intra-frame prediction applicable to V-PCC according to claim 7, characterized in that, The first-level loss uses a variant of Focal Loss when class imbalance exists.

10. The fast mode decision method for VVC intra-frame prediction applicable to V-PCC according to claim 1, characterized in that, The model input also includes: auxiliary features of geometric blocks and occupancy blocks; occupancy blocks are used to guide geometric video; when performing intra-frame coding on attribute video, auxiliary features of geometric blocks and occupancy blocks are used to participate in intra-frame mode decision-making, and geometric blocks and occupancy blocks guide attribute block coding.

Images