Point Cloud Compression via Occupancy Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current point cloud compression technologies, such as 3D native coding and 2D projection followed by traditional video coding, are limited in their ability to efficiently and flexibly compress large volumes of 3D data from sources like 3D scanners and VR/AR applications, particularly in handling dynamic scenes with millions of points per second.
Innovation Solution
Occupancy networks are used to implicitly represent 3D surfaces with continuous decision boundaries, enabling efficient and flexible compression by estimating the probability of occupancy for positions in 3D space, which is then used for entropy encoding, allowing for scalable and lossless or lossy compression of point clouds, meshes, or projected images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional video coding or 3D native coding is used for point cloud compression, then the transmission of large amounts of 3D data can be achieved, but the compression efficiency and flexibility are limited
Solution Approach 1:
The patent transforms the point cloud representation from discrete 3D coordinates to a continuous implicit neural function representation. This parameter transformation enables flexible resolution scaling by evaluating the same function at different densities, while achieving high compression by storing only function parameters rather than millions of point coordinates.
Solution Approach 2:
The patent replaces traditional mechanical compression algorithms (octree, voxel-based methods) with a machine learning-based implicit neural function approach. This substitution enables adaptive resolution control and continuous surface representation, providing both high compression ratios and flexible scalability without the rigidity of fixed-grid traditional methods.
2Measurement precision
If dense point clouds with thousands or millions of points are captured for high fidelity, then the representation accuracy is improved, but the data transmission burden increases significantly
Solution Approach 1:
The patent creates a compressed implicit function copy of the original dense point cloud that can generate high-fidelity 3D representations on-demand. Instead of transmitting millions of point coordinates, only the compact neural function parameters are transmitted, enabling accurate reconstruction with minimal data transmission.
Solution Approach 2:
The patent transitions from representing 3D space with discrete point clouds to representing it with a continuous implicit function in a higher-dimensional parameter space. This dimensional transformation allows the compact function to encode the same geometric information that would otherwise require millions of discrete points.
3Reliability
If every frame of dynamic 3D scenes is encoded as a unique dense point cloud, then the temporal fidelity is maintained, but the transmission rate requirements become unmanageable
Solution Approach 1:
The patent transforms dynamic point cloud sequences into sequences of implicit neural functions with evolving parameters. This transformation maintains temporal fidelity by capturing frame-to-frame variations in function parameters while achieving drastic compression, reducing transmission rates from millions of points per second to manageable parameter updates.
Data Source
AI summary
Occupancy networks enable efficient and flexible point cloud compression. In addition to the voxel-based representation, occupancy networks are able to handle points, meshes, or projected images of 3D objects, making them very flexible in terms of input signal representation. The probability of occupancy of positions is estimated using occupancy networks instead of sparse convolutional neural networks. A compression implementation using occupancy network enables scalability with infinite reconstruction resolution.