Autoregressive and Residual Index Convolution Model for Point Cloud Geometry Compression

TL;DR

This paper introduces a new point cloud compression model that uses autoregressive and residual index convolution techniques to improve compression performance and reduce resource usage.

Contribution

The novel contribution is the integration of autoregressive grouping and a distiller layer in a point cloud compression model.

Findings

The proposed model outperforms recent deep learning-based compression models in performance.

The approach reduces time and memory consumption compared to previous voxel convolution and attention-based methods.

The model demonstrates effectiveness on three datasets using autoregressive and residual techniques.

Abstract

This study introduces a hybrid point cloud compression method that transfers from octree-nodes to voxel occupancy estimation to find its lower-bound bitrate by using a Binary Arithmetic Range Coder. In previous attempts, we demonstrated that our entropy compression model based on index convolution achieves promising performance while maintaining low complexity. However, our previous model lacks an autoregressive approach, which is apparently indispensable to compete with the current state-of-the-art of compression performance. Therefore, we adapt an autoregressive grouping method that iteratively populates, explores, and estimates the occupancy of 1-bit voxel candidates in a more discrete fashion. Furthermore, we refactored our backbone architecture by adding a distiller layer on each convolution, forcing every hidden feature to contribute to the final output. Our proposed model extracts local features using lightweight 1D convolution applied in varied ordering and analyzes causal relationships by optimizing the cross-entropy. This approach efficiently replaces the voxel convolution techniques and attention models used in previous works, providing significant improvements in both time and memory consumption. The effectiveness of our model is demonstrated on three datasets, where it outperforms recent deep learning-based compression models in this field.

Figures