Dynamic Scene Reconstruction

Dynamic scene reconstruction estimates how a scene’s geometry, appearance, and motion change over time, enabling novel-view rendering, mapping, and simulation. Methods use neural radiance fields, 3D and 4D Gaussian splatting, meshes, or point-based models, guided by cues such as depth, optical flow, camera motion, object identity, and physical constraints. Work spans monocular and multi-view video, from deformable objects and human interactions to driving scenes, and increasingly uses feed-forward models alongside slower per-scene optimization. A central challenge is that realistic-looking renderings do not always imply accurate geometry or motion; methods must balance fidelity, physical plausibility, speed, and robustness to sparse views, occlusion, and long sequences.