ReconPhys: Reconstruct Appearance and Physical Attributes from Single Video
This addresses the problem of creating simulation-ready assets for robotics and graphics by eliminating expensive tuning or manual annotation, though it appears incremental as it builds on differentiable rendering and 3D Gaussian Splatting.
The paper tackles the challenge of reconstructing non-rigid objects with physical plausibility from single monocular videos, proposing ReconPhys as a feedforward framework that jointly learns physical attribute estimation and 3D Gaussian Splatting reconstruction without ground-truth physics labels. The method achieves 21.64 PSNR in future prediction (vs. 13.27 for baselines), reduces Chamfer Distance from 0.349 to 0.004, and enables fast inference in under 1 second compared to hours for existing methods.
Reconstructing non-rigid objects with physical plausibility remains a significant challenge. Existing approaches leverage differentiable rendering for per-scene optimization, recovering geometry and dynamics but requiring expensive tuning or manual annotation, which limits practicality and generalizability. To address this, we propose ReconPhys, the first feedforward framework that jointly learns physical attribute estimation and 3D Gaussian Splatting reconstruction from a single monocular video. Our method employs a dual-branch architecture trained via a self-supervised strategy, eliminating the need for ground-truth physics labels. Given a video sequence, ReconPhys simultaneously infers geometry, appearance, and physical attributes. Experiments on a large-scale synthetic dataset demonstrate superior performance: our method achieves 21.64 PSNR in future prediction compared to 13.27 by state-of-the-art optimization baselines, while reducing Chamfer Distance from 0.349 to 0.004. Crucially, ReconPhys enables fast inference (<1 second) versus hours required by existing methods, facilitating rapid generation of simulation-ready assets for robotics and graphics.