Accurate, dense, and robust multi-view stereopsis

Yasutaka Furukawa, Jean Ponce (2010)

Key points

Output: dense set of small, rectangular patches covering visible surfaces
Method is able to filter out features/patches that don't appear frequently
Hybrid approach, applicable to objects, crowded scenes, etc. using a flexible patch-based MVS algorithm
- Method is able to filter out features/patches (of outliers and obstacles) that don't appear frequently, allowing for application to crowded scenes
- Match, expand, filter: sparse set of matched keypoints/patches --> expand to nearby pixels to get dense set of patches --> visibility constraints + weak regularization to filter incorrect matches
Optional: convert into mesh, wihch can be further refined by enforcing photometric consistency
Algorithm doesn't require any initialization
Estimates local surface orientation and enforces photometric consistency
Accurate object model despite large/deep concavities or low texture
Lack of strong regularization helps in reconstructing deep concavities, but may cause problems in poor-texture regions
Also: because of lacking regularization, the method only reconstructs regions where there is reliable information (post-processing would be needed to fill the gaps)
Method is not real-time