As compositionality seems to be on the roadmap, I guess, sources for objects/features would be needed.
One problem which is typically solved via fancy algorithms in the CAx world is finding features in a CAD drawing, e.g. boreholes, chamfers or radii (rounded edges or other cavities). The drawings can be synthetically constructed in an OSS CAD.
Does that sound like a source of models?
The current Monty data set seems to mostly include convex shapes, and holes, of course, are concave features.
Some signs that this is a problem being tackled in the DL world too:
- Machining feature recognition based on deep neural networks to support tight integration with 3D CAD systems | Scientific Reports
- Feature recognition - Wikipedia
The simplicity of such models might also be a good testbed, since cad models can be generated programmatically, with fixed results, and thus fixed assertions, e.g. “3 radii of r=2, and 2 boreholes of r=3, and 1 borehole of r=5”. (also opening up a possibility of testing scale)
Although, even already tried scenes can probably be scripted and auto-evaluated as well.