Author: Salman
Year: 2010
- Dense 3D point cloud from multiview stereo, uses 2D delaunay triangulation
- according to A
comparison and evaluation of multi-view stereo reconstruction
algorithms, 4 class of methods:
- working on a 3D volume
- variationnal methods
- fusing multiple depth maps
- extract 3D points and fit a surface
- A triangular depth map is made for each images, then triangles lifted in 3D. The triangle soup is refined by combining visibility and photo consistency constraints (?)
- The Voronoi cell
$V(p_i)$ , associated to a point$p_i$ , is the region of points that are closer to$p_i$ than to all other sites in$P$ - Takes as input a set of calibrated images and a set of tracks
- Several steps:
- Merging close tracks
- Filters outliers by distance to nearest neighbours and low parallax
- Smooth the surface by polynomial fitting of the neighbours of each point and reprojection
- Detects contour by gaussian filter + canny, 2D triangulation over all tracks and 3D lifting
- filter triangle soup
- Filter triangle soup:
- visibility constraint: remove any triangle that is crossed by the ray of a 3D point
- if the angle between the normal and the ray of its vertex is over 80°
- big triangles that have normal shape have photoconsistency check i.e the triangle on every view should have a small NCC
- Reconstruction with Provably good sampling and meshing of surfaces but it needs a prior