Event

Occluded point cloud segmentation networks mislabel hidden components as the same category

1 reports · 1 sources

The paper submitted by Tsung-Lung Yang finds that when an entire region is occluded at test time, point cloud segmentation networks relabel hidden components into a single convergence category rather than simply missing them. Under a fixed synthetic layout, this category absorbs up to 70% of test points, and baseline mIoU drops to 0.15–0.27; under random layouts, the convergence category shifts to stairs, with predicted share 4.5–5.4 times the true value (three training seeds). The authors append 14 handcrafted geometric features (GPE) to each point, restoring mIoU to 0.54–0.92 and eliminating 64–100% of whole-object losses on PointNet++ and a PTv3-style Transformer under both layouts; on real S3DIS scans, whole-wall loss drops from 10 to 0. Occlusion augmentation is a strong alternative on mIoU, but the paper reports that under every setting it has more whole-object losses than GPE (e.g., 25–35 vs. 0).