Multi-View Reconstruction

Multi-view reconstruction is the pipeline that recovers both the 3D structure of a scene and the poses of the cameras that observed it, starting from a collection of $n\ge 2$ images. It extends two-view geometry from a single pair to many views and connects to feature-based alignment as its front end.

The pipeline is classically split in two:

1. Structure from Motion — recovers camera poses $(R_i,T_i)$ and a sparse 3D point cloud (one point per matched feature) by combining the essential matrix, triangulation, Perspective-n-Point, and bundle adjustment.
2. Multi-View Stereo — given the SfM-recovered poses, reconstructs a dense surface by solving a per-pixel depth search across multiple views, then fusing the depth maps into a mesh.

Notes

• 6.1. Structure from Motion (SfM) — incremental pipeline for sparse reconstruction
• 6.2. Multi-View Stereo (MVS) — dense reconstruction following SfM

See Also

• 5.3. Epipolar Geometry — the two-view geometry that seeds multi-view reconstruction
• 4. Feature-Based Alignment — the front end that supplies correspondences
• 7. Neural Fields — modern neural alternatives (NeRF, Gaussian Splatting) that sit downstream of SfM