CADSplat: Sparse-View 3D Gaussian Splatting Aided by CAD Models for Robust, Photorealistic Digital-Twin Reconstruction

arXiv cs.GR details CADSplat, a framework for reconstructing photorealistic, geometrically accurate digital twins from sparse posed images by regularizing 3D Gaussian Splatting with an explicit CAD shape prior. The paper was submitted on 16 Sep 2026 by Kristof Overdulve, Lode Jorissen, and Nick Michiels.

The practical hook is that CADSplat works with fewer than 15 wide-baseline views, and the method degrades gracefully even down to 3 images. It first matches segmented silhouettes against a CAD library to find a similar model and estimate camera-to-object poses, then anchors Gaussian primitives to that surface while jointly optimizing the splats, registration, and a smooth non-rigid deformation field.

For developers, that means a more robust route to digital-twin capture when full photogrammetry coverage is unrealistic. The authors say most of the quality gain comes from constraining the splats to a surface and moving them with a single deformation field, while the CAD prior helps most when views are sparse or the object is heavily self-occluded.

Beyond novel-view synthesis, the setup could feed markerless AR registration, per-image pose estimation, physics workflows, and part-label transfer from design assets to reconstructed objects. That makes it especially relevant for pipelines that need usable geometry from limited real-world capture rather than perfect scan conditions.