CN-121985129-A - Point cloud context neighborhood selection method for entropy model

CN121985129ACN 121985129 ACN121985129 ACN 121985129ACN-121985129-A

Abstract

The invention belongs to the technical field of learning type point cloud compression and entropy coding probability modeling, in particular to a point cloud context neighborhood selection method for an entropy model, which is applied to a learning type point cloud compression system, the system converts point cloud data into potential characteristics and quantifies the potential characteristics to obtain potential variables, the potential variables are divided into a first coding characteristic and a second coding characteristic according to a checkerboard grouping rule, the first coding feature is subjected to entropy coding in advance and recovered in advance at a decoding end, as decoded side information, the conditional probability of the second coding feature is jointly determined by the decoded first coding feature context neighborhood feature and the super priori side information, and the neighborhood feature of the second coding feature is searched in the feature space of the decoded first coding feature for decoding of the second coding feature. The invention reduces the uncertainty of conditional distribution by eliminating redundant/mismatched neighbors from the candidate pool, thereby reducing the coding bits.

Inventors

WANG DAYONG
ZHAO XIONG
Peng Kongqing
YUN YUANSHU

Assignees

重庆邮电大学

Dates

Publication Date: 20260505
Application Date: 20260211

Claims (9)

1. A point cloud context neighborhood selection method for an entropy model is applied to a learning point cloud compression system, the system transforms point cloud data into potential features and quantifies the potential variables, the potential variables are divided into first coding features and second coding features according to a checkerboard grouping rule, the first coding features are subjected to entropy coding in advance and restored in advance at a decoding end to serve as decoded side information, and the method is characterized in that conditional probability of the second coding features is determined by the decoded first coding feature context neighborhood features and super prior side information together, and a process for searching neighborhood features of the second coding features in a feature space decoded by the first coding features comprises the following steps: 101. For any position n to be coded of the second coding feature, searching to obtain a candidate neighbor set C n in the decoded first coding feature space according to a deterministic geometric neighbor rule, wherein the size of the candidate neighbor set is K cand ; 102. Selecting K kept elements from the candidate neighbor set as a reserved neighbor set S n , and performing T times of 1-swap replacement operation on the set to obtain a final neighbor set S' of a position n to be coded; 103. In the process of 1-swap replacement for the t time, selecting an element from the complement C n \S t-1 to replace the element in the reserved neighbor set S t-1 obtained by the previous 1-swap replacement, so that the local code rate change is minimum, wherein the complement C n \S t-1 is a set consisting of elements except the reserved subset S t-1 in the candidate neighbor set C n ; 104. And (3) taking elements in the final neighbor set S 'as neighbor contexts of a position n to be coded of the second coding feature, and inputting the elements into the neighbor set S' for conditional probability prediction and entropy coding/entropy decoding of the second coding feature.
2. The method according to claim 1, wherein the candidate neighbor set C n is obtained by searching according to deterministic geometric neighbor rules, i.e., K cand of geometric distances from the position n to be encoded are selected as candidate neighbor sets C n from all neighbors, and K kept of nearest neighbors to the position n to be encoded are selected when K kept elements are selected as reserved neighbor sets S n from the candidate neighbor sets.
3. The method for selecting a point cloud context neighborhood for an entropy model according to claim 1, wherein the local code rate change of the position n to be encoded in the process of the t-th 1-swap substitution is represented as: ; Wherein, the And S t represents a reserved neighbor set obtained by executing the t-1 st 1-swap replacement, namely a set obtained by replacing one element in the reserved neighbor set S t-1 obtained by executing the t-1 st 1-swap replacement with one element in the complement C n \S t-1 .
4. A method of selecting a point cloud context neighborhood for an entropy model according to claim 3, wherein the preserving subset S t is used as the context neighborhood and the set is used to encode the second encoding feature at the encoding cost of the position n to be encoded Expressed as: ; Wherein, the Conditional probability output for entropy model of learning point cloud compression system by preserving neighbor subset The super prior side information h is input to an entropy model of the learning point cloud compression system, and the entropy model outputs probabilities; representing a second encoding feature The code symbol at position n to be coded.
5. The method for selecting the point cloud context neighborhood for the entropy model according to claim 1, wherein the local code rate change of the position n to be coded is predicted by a perceptron network formed by a multi-layer perceptron, characteristic information of a reserved subset and a discarded subset, geometric relation characteristics of the reserved subset and the discarded subset respectively at the current position n to be coded and super prior side information are taken as inputs in the prediction, and the multi-layer perceptron outputs the local code rate change of the position n to be coded in the input state.
6. The method of claim 5, wherein parameters of the perceptron network are updated by minimizing a mean square error of the perceptron network predictions and training data when predicting the perceptron network.
7. A point cloud context neighborhood selection system for an entropy model, for implementing a point cloud context neighborhood selection method for an entropy model as claimed in claim 1, comprising an offline module and an online module, wherein: The off-line module is used for collecting samples for training, each sample comprises a feature and a label, the feature of the sample is composed of feature information of a reserved subset and a discarded subset of each coded position, geometric relation features of the reserved subset and the discarded subset with the current position n to be coded and super priori side information, and a neighbor set obtained through calculation in the steps 101-104 is used as the label; And the online module is used for training the perceptron network formed by the multi-layer perceptrons based on the sample data acquired by the offline module, and predicting the labels of the trained perceptrons based on the characteristics of the positions to be encoded.
8. A computer device, the device comprising: A memory; A processor; Wherein the memory stores computer-executable instructions; The processor executes computer-executable instructions stored in the memory to implement a point cloud context neighborhood selection method for an entropy model as claimed in any of claims 1-6.
9. A computer storage medium having stored therein computer executable instructions which when executed by a processor are for implementing a point cloud context neighborhood selection method for an entropy model as claimed in any of claims 1-6.

Description

Point cloud context neighborhood selection method for entropy model Technical Field The invention belongs to the technical field of learning type point cloud compression and entropy coding probability modeling, and is applied to an end-to-end point cloud compression coding and decoding system, in particular relates to a point cloud context neighborhood selection method for an entropy model. Background The point cloud is usually composed of geometric coordinates and attributes (such as colors), and has large data scale, sparsity, local irregularity and higher compression difficulty. Traditional point cloud compression standards (e.g., octree, prediction, transform, and entropy coding based schemes) often require a compromise between code rate and complexity in complex scenarios. In recent years, the learning type point cloud compression gradually adopts an end-to-end network and a probability model to carry out coding, and a typical framework comprises that an encoder/decoder obtains latent variable representation, and the super prior (hyperprior) and a context model are combined to predict the conditional probability of a latent variable symbol, and then bit output is realized through arithmetic coding. To improve probability prediction accuracy, existing methods typically build up a context Wen Linyu (e.g., based on KNN, octree neighborhood, or fixed geometry rules) that inputs the neighbor features of the decoded region into the context aggregation network. However, the neighborhood built by the fixed geometry rules has a natural limitation that the same geometric neighborhood is not equivalent to "useful" for target symbol probability prediction. Weak correlation, redundancy or mismatch neighbors tend to be mixed in the neighborhood, resulting in a "scattered" condition distribution, which increases the code rate. More importantly, in many parallelized entropy models (e.g., group/checkerboard modeling), neighborhood sources are often defined within decoded groups, and neighborhood construction is more easily "fixed by default" and lacks optimizable mechanisms. Disclosure of Invention Aiming at the problems that the prior learning type point cloud compression context neighborhood is directly determined by a fixed geometric rule and cannot be optimized, so that redundancy/mismatch neighbors are introduced and the prediction precision of an entropy model is reduced, the invention provides a point cloud context neighborhood selection method for an entropy model, which is applied to a learning type point cloud compression system, the system transforms point cloud data into potential features and quantifies the potential variables, the potential variables are divided into first coding features and second coding features according to a checkerboard grouping rule, the first coding features are subjected to entropy coding in advance and are recovered in advance at a decoding end to serve as decoded side information, the conditional probability of the second coding features is jointly determined by the decoded first coding feature context neighborhood features and the super priori side information, and the process for searching the neighborhood features of the second coding features in a feature space decoded by the first coding features comprises the following steps: 101. For any position n to be coded of the second coding feature, searching to obtain a candidate neighbor set C n in the decoded first coding feature space according to a deterministic geometric neighbor rule, wherein the size of the candidate neighbor set is K cand; 102. Selecting K kept elements from the candidate neighbor set as a reserved neighbor set S n, and performing T times of 1-swap replacement operation on the set to obtain a final neighbor set S' of a position n to be coded; 103. In the process of 1-swap replacement for the t time, selecting an element from the complement C n\St-1 to replace the element in the reserved neighbor set S t-1 obtained by the previous 1-swap replacement, so that the local code rate change is minimum, wherein the complement C n\St-1 is a set consisting of elements except the reserved subset S t-1 in the candidate neighbor set C n; 104. And (3) taking elements in the final neighbor set S 'as neighbor contexts of a position n to be coded of the second coding feature, and inputting the elements into the neighbor set S' for conditional probability prediction and entropy coding/entropy decoding of the second coding feature. The invention also provides a point cloud context neighborhood selection system for the entropy model, which is used for realizing a point cloud context neighborhood selection method for the entropy model, and comprises an offline module and an online module, wherein: The off-line module is used for collecting samples for training, each sample comprises a feature and a label, the feature of the sample is composed of feature information of a reserved subset and a discarded subset of each coded pos