Superpoint Transformer for 3D Scene Instance Segmentation

Most existing methods realize 3D instance segmentation by extending those models used for object detection or semantic segmentation. However, these non-straightforward suffer from two drawbacks: 1) Imprecise bounding boxes unsatisfactory predictions limit the performance of overall framework. 2) Existing method requires a time-consuming intermediate step aggregation. To address issues, this paper proposes novel end-to-end based on Superpoint Transformer, named as SPFormer. It groups potential features point clouds into superpoints, and directly predicts instances through query vectors without relying results The key in framework is decoder with transformers that can capture information superpoint cross-attention mechanism generate masks instances. Through bipartite matching masks, SPFormer implement network training aggregation step, which accelerates network. Extensive experiments ScanNetv2 S3DIS benchmarks verify our concise yet efficient. Notably, exceeds compared state-of-the-art 4.3% hidden test set terms mAP keeps fast inference speed (247ms per frame) simultaneously. Code available at https://github.com/sunjiahao1999/SPFormer.