Streaming 3D Reconstruction

Streaming 3D reconstruction estimates camera motion and scene geometry as frames arrive, producing point maps, meshes, or renderable Gaussian-splat scenes for robotics, mapping, and interactive viewing. Methods range from causal vision and geometry transformers and recurrent state models to SLAM pipelines with bundle adjustment, pose graphs, and incremental Gaussian or TSDF mapping; test-time training and external memories carry context across long sequences. To bound cost, systems compress or retrieve tokens, retain spatial anchors, and gate updates using novelty, confidence, or motion, while dynamic-scene methods add temporal deformation and semantic or object tracking. Evaluation on indoor RGB-D datasets such as 7-Scenes, ScanNet, and TUM, and outdoor driving sequences such as KITTI, measures pose drift, geometric accuracy, rendering quality, speed, and memory; the central challenge is preventing drift and forgetting without making updates too slow or unresponsive to scene changes.