Multi-View 3D Reconstruction

Multi-view 3D reconstruction combines images or video from different viewpoints to estimate camera poses, depth, surfaces, and sometimes appearance or motion. Classical structure-from-motion and multi-view stereo rely on feature matching, triangulation, and bundle adjustment, while neural methods use NeRFs, Gaussian splatting, and feed-forward geometry transformers to infer 3D from images. Current work also adds depth and vision-model priors, selects informative views, and handles difficult cases such as uncalibrated cameras, dynamic scenes, reflective materials, and wide-angle or satellite imagery. A central challenge is balancing accurate, complete geometry across sparse or inconsistent views with the speed, memory, and generalization needed for large-scale and real-time use.