Introducing paper, arXiv v1; engineering-data and MFCAD++ experiments are distinct.
What is annotated?
Human labels distinguish hole from non-hole B-Rep faces. The dataset contains 2,202 parts and 6,864 holes; 14,471 of 197,003 faces are hole faces.
Schütte, Heinert & Rottmann · Methods, Tables 1–3 and limitations (arXiv v1)
Rules and learned geometry
GWS searches boundary loops; MVR combines rendered views and topology. UV-Net and AAGNet learn face labels from B-Rep geometry.
Schütte, Heinert & Rottmann · Methods, Tables 1–3 and limitations (arXiv v1)
Detecting a complete hole
Connected predicted faces form hole candidates. Hole matching uses surface-area-weighted IoU ≥ 0.99; face-level metrics measure a different unit.
Schütte, Heinert & Rottmann · Methods, Tables 1–3 and limitations (arXiv v1)
Read both test settings
The study separately tests 200 selected MFCAD++ parts. Strong in-domain segmentation does not establish robustness to different CAD modelling conventions.
Schütte, Heinert & Rottmann · Methods, Tables 1–3 and limitations (arXiv v1)
Annotation and coverage limits
Multi-solid parts were excluded. Functional labels can omit geometrically valid holes, so some detections count as false positives despite finding real cavities.
Schütte, Heinert & Rottmann · Methods, Tables 1–3 and limitations (arXiv v1)
Published results: engineering CAD
The paper reports these values for its own dataset. Keep hole detection and face segmentation separate.
| Method | Hole recall | Hole precision | Hole F1 | Face recall | Face precision | Face F1 |
|---|---|---|---|---|---|---|
| AAGNet | 0.944 | 0.810 | 0.871 | 0.912 | 0.932 | 0.922 |
| UV-Net | 0.934 | 0.896 | 0.914 | 0.883 | 0.933 | 0.920 |
| GWS | 0.927 | 0.578 | 0.712 | 0.977 | 0.644 | 0.733 |
| MVR | 0.728 | 0.870 | 0.793 | 0.768 | 0.898 | 0.871 |
Values use the original 0–1 scale. Section 5.5 specifies area-weighted hole IoU ≥ 0.99. GWS and MVR are rules, not trained neural models. F1 values are reproduced as published, not recomputed.
Schütte, Heinert & Rottmann · Methods, Tables 1–3 and limitations (arXiv v1)
Published results: transfer to MFCAD++
The transfer evaluation uses 200 selected MFCAD++ parts. It is a different test setting from the preceding table.
| Method | Hole recall | Hole precision | Hole F1 | Face recall | Face precision | Face F1 |
|---|---|---|---|---|---|---|
| AAGNet | 0.013 | 0.357 | 0.026 | 0.023 | 0.433 | 0.044 |
| UV-Net | 0.468 | 0.933 | 0.623 | 0.381 | 0.957 | 0.545 |
| GWS | 0.889 | 0.805 | 0.845 | 0.983 | 0.768 | 0.871 |
| MVR | 0.617 | 0.987 | 0.760 | 0.577 | 0.987 | 0.729 |
Original 0–1 values; hole IoU threshold 0.99. Scores belong to this selected transfer set, not the full MFCAD++ collection.
Schütte, Heinert & Rottmann · Methods, Tables 1–3 and limitations (arXiv v1)