Structural Geometry Conditioning for Executable Mesh-to-CAD Reconstruction
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
An editable CAD model requires a valid construction program, whereas an observed mesh specifies only the resulting surface. This mismatch makes mesh-to-CAD reconstruction especially fragile for multi-part objects: dense surface samples describe shape but may obscure the profile boundaries and operation choices needed for a short executable program. We use an external part segmenter and study a geometry-conditioned program generator that augments surface tokens with curves extracted from each mesh part at inference, concatenates the two streams, and weights operation-name tokens during supervised training. Training uses reference B-Rep curves, so curve inputs differ between training and evaluation. A compiler prepares canonical training targets from source histories; a deterministic CAD executor converts predicted programs to B-Rep and STEP output. On the reported 12-case multi-part evaluation, the system has Chamfer distance 0.010054, 95th-percentile surface distance 0.032163, and 64^3 voxel IoU 0.580917. On all 1,000 validation part programs, operation weighting raises STEP export success for the all-curve model from 0.894 to 0.958 and reduces penalized mean CD from 0.0198 to 0.0162. Linear curves alone outperform the unfiltered all-curve input on several program metrics, revealing a curve-selection trade-off. These results support structural conditioning for executable reconstruction while leaving broader generalization to larger, fully audited evaluations.
- Primary area:
- foundation or frontier models, including LLMs