Point-cloud segmentation drops fixed pooling: COC learns a per-sample measure and lifts ScanNetV2 panoptic PQ from 71.2 to 72.0
Synopsis
Inside the unified OneFormer3D query decoder, regional feature aggregation previously relied on fixed statistics such as mean or max; COC replaces that step with a learned positive scalar weight per sample, adds contracted first- and second-order geometric descriptors, and evolves states bidirectionally along decoder depth. On ScanNetV2 instance mAP rises from 59.3 to 59.8, semantic mIoU from 76.6 to 77.4 and panoptic PQ from 71.2 to 72.0; on S3DIS Area-5 mIoU rises from 72.4 to 73.8. Gains vary by dataset and metric, and the authors note cross-paper numbers are reference points only.
Interpretation
A unified 3D indoor segmentation model can now assign each point a learned positive weight at the superpoint aggregation stage instead of collapsing a region into a mean or maximum. Earlier point-cloud pipelines summarized local features with fixed statistics or learned weights, leaving a gap between permutation-invariant pooling and geometry-aware integration. Under the same S3DIS Area-5 protocol, scatter mean pooling reaches 72.4 mIoU, scatter max 72.2, linear weighting 73.5, and LGDI 74.0.
The change replaces three information-transfer operations inside the OneFormer3D query decoder while the prediction heads stay unchanged. LGDI uses Softplus to produce a positive scalar measure per sample, normalizes it into weights, and combines a weighted content integral, a row-norm-contracted first-order Jacobian descriptor and paired cross-product second-order descriptors in a residual update. SAG-CA adds a permutation-invariant state bank that queries read alongside a local positive kernel read, and DSSA runs a bounded bidirectional recurrence only along the intrinsically ordered decoder-depth axis; experiments use fixed ScanNetV2 and S3DIS configurations on a single RTX 4090.
Perspective
Evaluation is confined to two indoor benchmarks, ScanNetV2 and S3DIS, covering semantic, instance and panoptic tasks, with all experiments run on a single RTX 4090. The authors list extension to large-scale outdoor and dynamic scenes such as autonomous-driving LiDAR as future work, so the current conclusions apply to indoor scenes.
The authors note that numerical gains are moderate for some individual metrics, that the tables do not establish statistical significance for every comparison, and that cross-paper numbers are reference points rather than strictly controlled same-run comparisons. The geometric invariance checks are operator-level tests with fixed random features and region assignments and do not imply invariance of the complete trained network; the S3DIS training configuration sets rotation range to zero, so its scores do not constitute a rotation-robustness evaluation.