PillarCoder: An Efficient and Lightweight Attention Network With Pillarization for Point Cloud Learning

Existing point cloud learning networks based on the attention mechanism suffer from high computational and storage resources, as the generation of attention maps and their usage are with respect to each point. To address this problem, we propose an efficient and lightweight point cloud learning network termed PillarCoder, which performs the generation and usage of attention maps on a finite number of pillars through the pillarization, rather than on each point. Specifically, the pillar extractor is first applied to extract shape features inscribed by points within each vertical pillar, and the semantic embedding and contextual position encoding are designed to enhance its capability in terms of semantic capture and positional information retention, respectively. Second, the convolution aggregator is proposed to aggregate the global feature of a point cloud from the dense pillar-level features with fewer resource requirements. Then, the information retrieval module is used to match each point feature with its corresponding pillar feature and global feature, making each point perceive both local and global spatial information. Experimental results show that PillarCoder is able to maintain competitive results with faster inference and lower requirements on computational and storage resources, compared to existing attention networks.