4D Reconstruction

4D reconstruction studies how to recover both 3D structure and its changes over time, from whole scenes and human interactions to organs and plants. Methods use representations such as dynamic Gaussian splats, neural fields, and deformable meshes, with cues from multiple views, optical flow, depth, temporal models, and physical or anatomical priors. Work ranges from per-scene optimization to fast feed-forward transformers and generative models, with benchmarks increasingly testing long videos, sparse views, occlusion, and motion tracking. A central challenge is producing geometrically and temporally consistent, physically plausible reconstructions from incomplete observations without sacrificing detail, speed, or generalization.