STL to STEP accuracy benchmark
Last updated by the stl2step team
On 200 CAD parts it never saw in training, stl2step recovered 97.9% of the original faces, covering 99.2% of the surface area, and gave 99.0% of them the right surface type. On 169 of the 200 parts every single face came back.
- 97.9%
- of 8,380 faces recovered
- 99.2%
- of the surface area covered by recovered faces
- 99.0%
- of faces with the correct surface type
Method
- Parts: the first 200 usable closed parts of chunk 0009 of the ABC dataset (one million CAD models from Onshape). No part of that chunk was used to train the model.
- Input: each part as a closed, welded triangle mesh, median 730 triangles, up to 38,784. Only the mesh is given to stl2step.
- Reference: the faces of the original B-Rep, labeled by what each face is geometrically (a cylinder split into two half-faces by the file counts as one cylinder).
- Scoring: predicted and original faces are matched one to one; a face is recovered when the overlap (IoU) of their triangles is at least 0.9. Reported by count and by area.
- System: the same pipeline that runs on stl2step.ai, on CPU, with no manual input.
Results by part complexity
| Faces in the original part | Parts | Faces recovered |
|---|---|---|
| 1–9 | 48 | 100.0% |
| 10–29 | 89 | 98.3% |
| 30–99 | 40 | 96.9% |
| 100 or more | 23 | 93.4% |
Simple parts come back complete; the hardest parts, with 100 faces or more, still keep 93.4% of their faces. 184 of the 200 parts have at least 90% of their faces recovered.
What the network adds
The same pipeline without the neural network, fitting surfaces from geometry alone, recovers 94.6% of the faces and 96.7% of the area. The network lifts that to 97.9% and 99.2%: it decides which triangles belong together where pure geometry is ambiguous, such as fillets meeting cylinders and faces split by tessellation.
A real part
The grip on the home page is not from ABC: 8,672 triangles become 364 faces with a mean distance of 0.003 mm between the mesh and the result on a 106 mm part.
Data
Per-part results (ABC part ID, faces, triangles, fraction of faces recovered): abc-200.csv, licensed CC BY 4.0. Cite as "stl2step face-recovery benchmark, https://stl2step.ai/benchmark".