Hierarchical Attention Networks for Lossless Point Cloud Attribute Compression

TL;DR

The paper tackles lossless point cloud attribute compression by introducing HA-PCAC, a hierarchical attention context model guided by a fast Hilbert-based Level of Detail (LoD) and level-wise autoregressive coding. The method partitions the point cloud into slices and refinement levels, enabling parallel processing within a level while a two-stage attention network extracts broad context and fine details; residual learning and normalization promote scale-invariant performance. The key contributions include (i) a fast LoD construction with parallelizable context groups, (ii) a residual, two-stage hierarchical attention model that predicts Laplace distribution parameters for attributes, and (iii) extensive experiments showing superior coding efficiency and runtime over the state-of-the-art G-PCC across color and reflectance attributes on diverse datasets. The approach improves generalization across varying densities and scales, offering practical impact for efficient lossless compression in real-world 3D scanning and autonomous systems.

Abstract

In this paper, we propose a deep hierarchical attention context model for lossless attribute compression of point clouds, leveraging a multi-resolution spatial structure and residual learning. A simple and effective Level of Detail (LoD) structure is introduced to yield a coarse-to-fine representation. To enhance efficiency, points within the same refinement level are encoded in parallel, sharing a common context point group. By hierarchically aggregating information from neighboring points, our attention model learns contextual dependencies across varying scales and densities, enabling comprehensive feature extraction. We also adopt normalization for position coordinates and attributes to achieve scale-invariant compression. Additionally, we segment the point cloud into multiple slices to facilitate parallel processing, further optimizing time complexity. Experimental results demonstrate that the proposed method offers better coding performance than the latest G-PCC for color and reflectance attributes while maintaining more efficient encoding and decoding runtimes.

Figures (5)

• Figure 1: The point cloud is first partitioned into multiple slices. The points in each slice are then organized into several target groups with resolutions ranging from coarse to fine using our fast Level of Detail (LoD) construction. A deep hierarchical attention context model is employed to estimate the probability distribution of the attributes.
• Figure 2: Illustration of fast LoD construction.
• Figure 2: Comparisons of the proposed HA-PCAC and G-PCC-PLT with varying downsample step sizes and different reflectance quantization precisions
• Figure 3: Overall Architecture.
• Figure 4: Convergence of attribute values vs. residuals as inputs under identical training conditions on ScanNet.