Feed-Forward 3D Reconstruction

Feed-forward 3D reconstruction studies how to infer 3D shape, camera poses, and appearance from one image, sparse views, or video in a single pass or a small number of learned updates. Common approaches use visual geometry transformers, multi-view stereo features, and 3D Gaussian splatting, with pixel-aligned predictions alongside voxel-, point-, and token-based scene representations; some models jointly estimate unknown camera poses or maintain compact memory for streaming video. Work spans objects, indoor and outdoor scenes, dynamic driving environments, and human avatars, and is evaluated using novel-view rendering, depth and point-map accuracy, and camera-pose benchmarks. The central challenge is balancing fast, generalizable inference with globally consistent geometry, especially for sparse or unposed inputs, long sequences, and unseen viewpoints where scale drift or plausible but unsupported content can arise.