ASRL: Adaptive Sparse Representation Learning for LiDAR Point Cloud Geometry Compression

Octree-structured context entropy models with Transformers have demonstrated strong performance in LiDAR point cloud geometry compression (PCGC) by capturing long-range dependencies, which are crucial for superior rate-distortion (R-D) performance. While existing efficient attention mechanisms reduce computational cost, they often process irrelevant nodes, introducing redundancy and noise. In this work, we propose an Adaptive Sparse Representation Learning (ASRL) framework to mitigate the noisy interactions of irrelevant nodes and remove feature redundancy in cross-spatial-channel context for LiDAR PCGC. Specifically, we employ a sliding context window to process the resulting 1D sequence, which is then fed into multiple Adaptive Sparse Transformer (AST) blocks to enable efficient contextual feature representation learning. Each AST block contains two components: Adaptively-enhanced Sparse Spatial Interaction (ASSI), which combines dense and sparse attention branches through adaptive weighting to model global-local interactions, and Locally-enhanced Harmonized Channel Mixer (LHCM), which aggregates features across refinement scales and harmonizes channel-wise information via complementary interactions. Experiments on large-scale real-world LiDAR point cloud datasets with varying sparsity levels including SemanticKITTI (64-beam) and Ford (32-beam) show that ASRL achieves state-of-the-art compression performance and strong generalization across beam configurations.