Inputs → predictions → measured scores.
Leave one CAD design out: all scans sharing the held-out CAD are excluded from that fold’s training. Test results below use saved predictions on the same frozen CAD-node grid across sizes.
“Target” means the saved weak label: NG defect regions projected from CSV/scan evidence, or healthy targets inferred from production OK status. These are not independently reviewed ground truth. Scans and CSVs provide QA/targets; predictive inputs are CAD geometry and Excel process conditions.
Preprocessing types and input sizes
Images include 17 geometry/field channels and 17 validity channels. Unavailable physical fields retain their validity masks. Surface voxels have four channels. Graph nodes contain six geometry features plus six validity flags; graph edges have four features. UV grids use eight geometry/mask channels, broadcast process conditions, and face adjacency. Preview sampling changes only the displayed image.
Model × input size × LOO scores
Each row combines its actual preprocessing size and tensor dimensions with held-out test scores. BCE and AP show scan-macro mean, min, max and valid-case count. AP excludes cases with no positives. IoU is evaluated at the saved cutoff 0.5; empty unions remain undefined. Bold marks the best mean in each score column, including ties. Partial sizes show completed models only; pending and unavailable models stay explicit.
Best and worst test cases · every model
Best/worst selection is descriptive after evaluation, not a validation set or a tuning rule. NG is selected initially so easy healthy cases do not dominate the ranking. Ties are shown explicitly. Click a case to inspect its nodes and metric calculations.
One held-out case: actual input | target | model output
Target labels · weak
Red = target defect · blue = assumed healthy · grey = outside scoring mask. Click a node in this or either output map to inspect its probability and BCE.
Predicted defect probability
Target vs predicted labels
Green TP · purple FP · orange FN · blue TN · grey masked. Uses the interactive cutoff below.
Excel process conditions that also enter this model
How IoU, BCE and AP are measured
Moving this cutoff changes predicted binary labels and IoU. BCE uses the original probabilities; AP integrates ranking across all distinct score cutoffs. Neither changes when this slider moves. This is an explanation tool; it does not overwrite the saved test scores or select a production cutoff.
IoU · overlap of defect nodes
Target-positive set = saved target ≥ 0.5. Predicted-positive set = probability ≥ chosen cutoff. IoU = TP / (TP + FP + FN). True negatives do not enter the union.
BCE · probability penalty at every scored node
BCE = mean[−y ln(p) − (1−y) ln(1−p)]. The saved scorer clips p to [10⁻⁷, 1−10⁻⁷]. Confident wrong probabilities receive a large penalty. Every eligible node contributes; each scan then receives one vote in the macro mean.
AP · ranking across every cutoff
Sort probabilities from high to low. At each distinct score, include the whole tied group and compute precision and recall. AP = Σ (recall increase × precision). The shaded step rectangles show that sum; this is not trapezoidal PR area.
Actual ranked test nodes
Left = highest probability. Top bars show predicted probabilities; red markers indicate actual target-positive nodes. A useful ranking moves these markers toward the left. Masked nodes are excluded from every calculation.
Every eligible test case · selected profile
All 6 NG and 5 OK cases are listed for each completed model, irrespective of the NG/OK filter used for best/worst illustrations. Only finite metrics enter means and extrema; excluded and unavailable cases are not silently assigned zero. Use the source JSON links to inspect the original predictions.