Point Cloud Geometry Coding with Dynamic Predictors for Sparse LiDAR Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing point cloud codecs fail to efficiently combine simplicity, low latency, and high compression performance, particularly for sparse geometry data captured by spinning sensor heads, which are crucial for applications like autonomous driving.
Innovation Solution
A dynamic list of prediction data is used to derive candidate predictors for encoding and decoding point clouds, updating based on residual radius data to adapt to changing object distances, improving prediction efficiency and reducing bitrate.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If existing point cloud codecs are used, then encoding and decoding can be performed, but compression performance is insufficient and latency is high for sparse geometry data from spinning sensor heads
Solution Approach 1:
The patent implements dynamic predictor selection and updating mechanisms where the encoder and decoder maintain lists of candidate predictors that are dynamically updated based on previously decoded points. This allows the system to adapt to changing object distances and geometries in real-time, improving compression performance while maintaining low latency through efficient dynamic updates rather than static pre-computed predictors
Solution Approach 2:
The patent employs feedback mechanisms where decoded point information is fed back into the prediction process. The encoder and decoder use previously decoded points to generate updated candidate predictors, creating a feedback loop that continuously improves prediction accuracy. This feedback-based approach enables better compression performance while managing computational latency through iterative refinement
2Loss of time
If simple encoding methods are used, then latency is reduced, but compression performance deteriorates
Solution Approach 1:
The patent pre-generates lists of candidate predictors that are prepared in advance during the encoding/decoding process. These candidate lists are built using previously decoded information and are ready for immediate selection, eliminating the need for complex real-time optimization while maintaining high compression performance through pre-computed prediction options
Solution Approach 2:
The system uses dynamic candidate list management where predictors are added, removed, or updated based on the decoded geometry data. This dynamic adaptation allows simple per-point selection operations to achieve high compression performance by leveraging the most relevant predictors from the dynamically maintained candidate lists
3Loss of information
If high compression performance is achieved through complex prediction methods, then bitrate is reduced, but device complexity increases
Solution Approach 1:
The patent segments the prediction process into distinct stages: candidate predictor generation, candidate list management, and predictor selection. Each stage handles a specific aspect of the prediction task, allowing complex compression performance to be achieved through modular, manageable operations that reduce overall encoder complexity through structured segmentation
Solution Approach 2:
The encoder and decoder autonomously maintain and update their own candidate predictor lists using previously decoded information from the bitstream. This self-service mechanism eliminates the need for external predictor transmission or complex coordination, reducing device complexity while achieving high compression performance through self-contained prediction systems
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
There is provided methods and apparatus of encoding/decoding a point cloud Into/from a bitstream of encoded point cloud data representing a physical object. The methods determine a dynamic list of at least one prediction data used to derive at least one candidate predictor used to encode geometry data of points of the point cloud. The list of at least one prediction data is dynamic because during the encoding of points, prediction data are updated based on residuals radius of decoded points.