9 methods evaluated across 4 learning paradigms on the Industrial3D test set (Areas 6 + 12, 84.9M points). Metric: mean IoU (mIoU) over 12 classes via vote-based evaluation.
| # | Method | Paradigm | Labels | mIoU (%) | OA (%) | Year | Code / Status |
|---|---|---|---|---|---|---|---|
| Loading results… | |||||||
Click column headers to sort. Results verified by the Industrial3D team. OA = overall accuracy. — = not reported.
Vote-based per-class IoU for methods with full results available. Head classes dominate performance; tail classes (Reducer, Strainer, Pump, Elbow, Tee) remain near zero for most methods.
| Loading… |
Cyan = best per class among listed methods. 0.00 = zero IoU on that class. Only vote-based (smooth test) results shown.
Best supervised method (Boundary-CB): 55.74% mIoU
Best zero-shot foundation model (Point-SAM One-vs-Rest): 15.79% mIoU
A 39.95 percentage-point gap quantifies the unresolved industrial domain transfer challenge. Even the oracle Point-SAM (21.08%) — given GT mask proposals — trails the best supervised method by 34.66 pp. This gap arises from the dual crisis: foundation models trained on architectural/outdoor data are ill-equipped to distinguish geometrically ambiguous MEP fittings under extreme class imbalance.
Remarkably, SQN with only 0.1% labels (44.29%) outperforms fully-supervised RandLA-Net (39.83%), suggesting that the inductive bias of sparse-label training may be better suited to this domain than dense supervision with vanilla architectures.
Submit your method's results to the leaderboard. We welcome results from all paradigms, including few-shot, semi-supervised, and multi-modal approaches.
Submit Results →