EFITFormer: Enhanced Feature Interaction Transformer for 3D Point Cloud Understanding

The remarkable performance of Transformer architectures in the field of Natural Language Processing (NLP) has been extended to 2D computer vision, achieving tremendous success in various image/video-based tasks. And their unique property of permutation invariance makes them highly suitable for dealing with unstructured 3D point clouds. Therefore, in this paper, we introduce an Enhanced Feature Interaction Transformer (EFITFormer) for 3D point cloud understanding. Our architecture includes three imperative modules, namely a Point Pyramid Transformer (PPTransformer), a Multi-Level Transformer (MLTransformer) model, and a Mult-Scale Transformer (MSTransformer) module. Specifically, we present PP-Transformer to capture features with varying resolutions to enhance the power of learned features. Then, the MLTransformer is constructed to model the feature interactions across different levels, encoding geometric-semantic information. Finally, we design the MSTransformer to obtain long-range dependencies among different scales. Extensive evaluations on public benchmark datasets demonstrate the effectiveness and competitive performance of our methods in 3D shape classification, part segmentation, and semantic segmentation tasks.