Weak-label diagnostic only. Both samples are NG 찍힘, registration and defect extent are unvalidated, and local heater projection is invalid. Type is therefore a constant diagnostic label, never a learnable classifier.
image_2d prepared input · voxel_3d prepared input · point_6000 prepared input · graph_6000 prepared input
| representation | nodes/run | best loss | 18h01 MAE mm | 18h01 RMSE mm | CAD-correspondent Chamfer mm | registered-scan Chamfer mm | 18h01 weak AP |
|---|---|---|---|---|---|---|---|
| image_2d | 6000 | 0.3411 | 4.516 | 5.829 | 4.209 | 12.536 | 1.0 |
| voxel_3d | 1712 | 0.4047 | 5.999 | 7.865 | 5.591 | 15.891 | 0.6666666666666666 |
| point_6000 | 6000 | 0.4608 | 5.861 | 7.657 | 5.262 | 14.289 | 0.005 |
| graph_6000 | 5925 | 0.4284 | 5.851 | 7.846 | 4.634 | 15.277 | 0.479335311153493 |
image_2d attribution (JSON) · voxel_3d attribution (JSON) · point_6000 attribution (JSON) · graph_6000 attribution (JSON)
Paper-inspired three-level kd-tree Graph U-Net inside the conditional DDPM. It is a weak-label, two-run fit and is not a validation result.
| 18h01 | MAE mm ↓ | CAD Chamfer mm ↓ | weak AP ↑ | IoU@0.5 ↑ | uncertainty mm ↓ |
|---|---|---|---|---|---|
| graph_6000 | 5.851 | 4.634 | 0.479 | 0.0873 | 8.45 |
| graph_unet_6000 | 8.400 | 6.176 | 0.268 | 0.0019 | 14.11 |
same-view graph baseline vs TAG U-Net · metric comparison · kd-tree hierarchy · hierarchy JSON · attribution · full run report
