Iterative Tree-structured Correlation Diffusion for Hierarchical Dynamic Point Cloud Compression

Overview

With the rapid advancement of 3D acquisition technology, the growing volume of dynamic point cloud data has created an increasing demand for efficient compression methods. Octree, known for its variable resolution capabilities, has been widely adopted in point cloud compression. However, in dynamic point cloud compression scenarios, establishing correlations between reference octree nodes and encoding octree nodes presents significant challenges. Furthermore, effectively leveraging the spatial and temporal information embedded in octree sequences also faces several difficulties.

To address these issues, this project proposes a tree-structured iterative correlation diffusion (TICD) method to establish robust correlation links between the reference octree and the encoding octree. The approach involves a top-down search process to propagate node correlations progressively from coarse levels (the root node) to finer levels (non-leaf nodes). This enables more effective reference region identification and context utilization.

Building on this, we design a hierarchical predictive coding network (HSTC-Net) that integrates both temporal and spatial information while employing a hierarchical encoding strategy. This allows for batch-parallel processing and accurate occupancy predictions, significantly enhancing coding efficiency. The experimental results demonstrate that the proposed approach outperforms existing methods in compression efficiency and reconstruction accuracy, providing an innovative solution for dynamic point cloud storage and transmission.

Pipeline Illustration

(Pipline of TICD)

(Pipline of HSTC-Net)

Installation & Usage

1. Download Pre-trained Models

Before running the code, download the two pre-trained models from the provided link and place them in the save/ directory.

2. Setup Environment

Ensure you have Python 3.10 installed. Then, create a Conda environment and install the dependencies:

conda env create -f environment.yaml
conda activate your_env_name

3. Build Required Binaries

Compile the necessary binary files by running:

python setup.py build

4. Set Execution Permissions

Grant execution permission to utils/pc_error_d for distortion testing:

chmod +x utils/pc_error_d

5. Run Evaluation

Use the following command to perform encoding and decoding on the KITTI dataset:

python eval.py {your_kitti_path}/data_odometry_velodyne/dataset/sequences --decode

This script will compress and decompress the first 30 frames of all sequences in the dataset. If you need to modify the number of frames or sequences processed, you can edit line 67 in eval.py.

Some results used in the paper are available in the evaluation/kitti directory for reference. However, note that these results were generated using an earlier model and may exhibit slight differences from the latest test results.

License

This project is released under the BSD 2-Clause License. See the LICENSE file for details.