DALD-PCAC: Density-Adaptive Learning Descriptor for Point Cloud Lossless Attribute Compression

TL;DR

The paper tackles lossless point cloud attribute compression across varying densities by introducing a Levels of Detail (LoD) framework and a Density-Adaptive Learning Descriptor (DALD). It employs a permutation-invariant Transformer-based deep entropy model, guided by base-layer priors, and a block-partitioning strategy to manage variance and enable parallel decoding. The approach demonstrates strong bitrate reductions and robustness to density changes on LiDAR and object point clouds, while maintaining low computational complexity and fast decoding. The work points to practical benefits for real-world deployments and suggests extending the framework to lossy coding and broader point-cloud processing tasks.

Abstract

Recently, deep learning has significantly advanced the performance of point cloud geometry compression. However, the learning-based lossless attribute compression of point clouds with varying densities is under-explored. In this paper, we develop a learning-based framework, namely DALD-PCAC that leverages Levels of Detail (LoD) to tailor for point cloud lossless attribute compression. We develop a point-wise attention model using a permutation-invariant Transformer to tackle the challenges of sparsity and irregularity of point clouds during context modeling. We also propose a Density-Adaptive Learning Descriptor (DALD) capable of capturing structure and correlations among points across a large range of neighbors. In addition, we develop a prior-guided block partitioning to reduce the attribute variance within blocks and enhance the performance. Experiments on LiDAR and object point clouds show that DALD-PCAC achieves the state-of-the-art performance on most data. Our method boosts the compression performance and is robust to the varying densities of point clouds. Moreover, it guarantees a good trade-off between performance and complexity, exhibiting great potential in real-world applications. The source code is available at https://github.com/zb12138/DALD_PCAC.

Figures (11)

• Figure 1: The average of the number of neighbors (NN) that are captured by a $5\times5\times5$ convolution kernel for different point clouds. The number of neighbors is less than 10 points in most sparse point clouds.
• Figure 2: LoD architecture of our method. Input point cloud is sampled into Base Layer and Inference Layer which are coded separately. ${R_1,...,R_L}$ represents refinement layers. "DALD" represents Density-Adaptive Learning Descriptor (Sec. \ref{['DALD']}). Blocks are compressed by the Deep Entropy Model (Sec. \ref{['DALD']}) in parallel.
• Figure 3: Flowchart of the proposed attribute compression framework. Dashed lines represent providing context.
• Figure 4: LoD generation example. Points of each color form a refinement layer. In the Inference Layer, intra-prediction (blue arrow) is disabled.
• Figure 5: Comparison of block partitioning methods based on KD-tree and the proposed Base Layer priors.