LCPFormer: Towards Effective 3D Point Cloud Analysis via Local Context Propagation in Transformers

Transformer with its underlying attention mechanism and the ability tocapture long-range dependencies makes it become a natural choice for unorderedpoint cloud data. However, separated local regions from the general samplingarchitecture corrupt the structural information of the instances, and theinherent relationships between adjacent local regions lack exploration, whilelocal structural information is crucial in a transformer-based 3D point cloudmodel. Therefore, in this paper, we propose a novel module named Local ContextPropagation (LCP) to exploit the message passing between neighboring localregions and make their representations more informative and discriminative.More specifically, we use the overlap points of adjacent local regions (whichstatistically show to be prevalent) as intermediaries, then re-weight thefeatures of these shared points from different local regions before passingthem to the next layers. Inserting the LCP module between two transformerlayers results in a significant improvement in network expressiveness. Finally,we design a flexible LCPFormer architecture equipped with the LCP module. Theproposed method is applicable to different tasks and outperforms varioustransformer-based methods in benchmarks including 3D shape classification anddense prediction tasks such as 3D object detection and semantic segmentation.Code will be released for reproduction.