Definition
Multi-view stereo (MVS) extends stereo matching to views with known camera poses (e.g. from Structure from Motion) and produces a dense reconstruction (depth maps, points or a surface). In the reconstruction pipeline it comes after SfM: sparse model → MVS → dense model.
| Method | Idea |
|---|---|
| Voxel coloring (Seitz et al., 1999) | project each voxel into the images where it is visible, keep it if the colors are consistent (low variance), carve it otherwise; back-to-front visibility order |
| Plane-sweep stereo | project all views onto fronto-parallel planes at different depths; per pixel pick the depth with the best photo-consistency |
| Multi-view stereopsis (Furukawa et al., 2007) | match patches along epipolar lines, triangulate, expand and filter, reconstruct the surface (Poisson) |
| MVSNet (Yao et al., 2018) | learned features, plane sweep with differentiable homography, variance cost volume, 3D CNN, depth map |
Trap
MVS needs the camera poses. SfM gives poses + a sparse point cloud; MVS makes it dense. Newer methods (e.g. VGGT, DUSt3R) predict poses and geometry together from unposed images.
Appears in
- Lecture 4, Multi-View Stereo Matching
- Lecture 4, MVSNet
- Lecture 3, Reconstruction pipeline: sparse → dense
- Lecture 1, course overview: Building Rome in a Day
- Lecture 9, MVS with MASt3R: triangulation beats regression