🤖 AI Summary
This work addresses the supervision inconsistency between scene-level annotations and point-wise prediction targets in weakly supervised point cloud semantic segmentation. To this end, we propose a label propagation framework based on unsupervised oversegmentation and bipartite graph matching. First, unsupervised clustering (e.g., K-means) partitions the point cloud into semantically coherent superpoints (clusters). Second, a bipartite graph is constructed between clusters and scene-level labels, and the Hungarian algorithm is employed to achieve precise, sparse assignment of scene labels to the most relevant clusters. Third, a “conservative pseudo-labeling” strategy is introduced: only high-confidence clusters receive propagated labels, while remaining points leverage intrinsic cluster structure for self-driven learning. To our knowledge, this is the first work to incorporate bipartite graph matching into weakly supervised point cloud segmentation, effectively mitigating label noise and over-propagation. Our method achieves 62.3% mIoU on ScanNet and S3DIS—surpassing all existing weakly supervised approaches and approaching fully supervised state-of-the-art performance.
📝 Abstract
We propose a weakly supervised semantic segmentation method for point clouds that predicts"per-point"labels from just"whole-scene"annotations while achieving the performance of recent fully supervised approaches. Our core idea is to propagate the scene-level labels to each point in the point cloud by creating pseudo labels in a conservative way. Specifically, we over-segment point cloud features via unsupervised clustering and associate scene-level labels with clusters through bipartite matching, thus propagating scene labels only to the most relevant clusters, leaving the rest to be guided solely via unsupervised clustering. We empirically demonstrate that over-segmentation and bipartite assignment plays a crucial role. We evaluate our method on ScanNet and S3DIS datasets, outperforming state of the art, and demonstrate that we can achieve results comparable to fully supervised methods.
