PointMixer: MLP-Mixer for Point Cloud Understanding

MLP-Mixer has newly appeared as a new challenger against the realm of CNNs and Transformer. Despite its simplicity compared to Transformer, concept channel-mixing MLPs token-mixing achieves noticeable performance in image recognition tasks. Unlike images, point clouds are inherently sparse, unordered irregular, which limits direct use for cloud understanding. To overcome these limitations, we propose PointMixer, universal set operator that facilitates information sharing among unstructured 3D cloud. By simply replacing with Softmax function, PointMixer can “mix” features within/between sets. doing so, be broadly used intra-set, inter-set, hierarchical-set mixing. We demonstrate various channel-wise feature aggregation numerous sets is better than self-attention layers or dense token-wise interaction view parameter efficiency accuracy. Extensive experiments show competitive superior semantic segmentation, classification, reconstruction Transformer-based methods.