Reproduce “CADSplat: Sparse-View 3D Gaussian Splatting Aided by CAD Models for Robust, Photorealistic Digital-Twin Reconstruction” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kristof Overdulve, Lode Jorissen, Nick Michiels
CADSplat reconstructs objects from sparse photographs by retrieving a CAD model from silhouettes, aligning it with relative camera calibration, and anchoring 3D Gaussian splats to a deformable surface. It fits silhouettes before learning appearance. On six objects, CAD initialization improves three-view masked PSNR and IoU over a sphere, but slightly trails it with 9–13 views; these tests reuse calibration.
Key insights
- The CAD model acts as a structural prior rather than rigid ground truth: the deformation field accommodates design variations, tolerances, and imperfect model matches while preserving surface coherence.
- Surface-linked Gaussians avoid the free-space drift common in unconstrained 3DGS, and silhouette consensus provides a practical route to camera-to-CAD registration.
- The benefit of CAD initialization grows as views become sparser: with three views, it improves sphere initialization by 1.4 dB PSNR and 4.9 percentage points in silhouette IoU.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.