PointInst3D: Segmenting 3D Instances by Points

The current state-of-the-art methods in 3D instance segmentation typically involve a clustering step, despite the tendency towards heuristics, greedy algorithms, and lack of robustness to changes data statistics. In contrast, we propose fully-convolutional point cloud method that works per-point prediction fashion. doing so it avoids challenges clustering-based face: introducing dependencies among different tasks model. We find key its success is assigning suitable target each sampled point. Instead commonly used static or distance-based assignment strategies, use an Optimal Transport approach optimally assign masks points according dynamic matching costs. Our achieves promising results on both ScanNet S3DIS benchmarks. proposed removes inter-task thus represents simpler more flexible framework than other competing methods, while achieving improved accuracy.