Monocular 3D Reconstruction

Monocular 3D reconstruction turns images or video from one camera into estimates of shape, depth, camera motion, and sometimes changing objects or people. Methods range from structure-from-motion and SLAM to learned depth and point-map models, neural radiance fields, and 3D Gaussian splatting, often strengthened with geometric, semantic, or physical priors. Current work increasingly uses pretrained vision models and feed-forward systems for faster reconstruction, while specialized methods address people, hands, interactions, medical scenes, and long videos. A central challenge remains that one view does not uniquely determine 3D scale or hidden structure, especially when the camera and scene both move; accuracy, temporal consistency, generalization, and runtime can trade off.