Fitting of geometric transformations, 3D triangulation
Downloads
The Matlab code used is here.
This is one of the simpler transformations.
House Residual: 8.4665
Library Residual: 7.7973
The selected consensus points are shown in yellow. The initial points are marked with a square.
House Residual: 13.7116 inliers: 27
Library Residual: 24.5297 inliers: 42
Normalizing the points used seems to have a large effect. Note that for some reason my Library case actually had more error when normalized.
House - unnormalized Residual: 8.8149
House - normalized Residual: 5.9939
Library - unnormalized Residual: 1.8849
Library - normalized Residual: 2.436
The selected consensus points are shown in yellow. The initial points are marked with a square.
House Residual: 13.7116 inliers: 27
Library Residual: 24.5297 inliers: 42
The red lines are segments of the epipolar lines. The blue points are the original matches in the image. A purple line represents the residual from the actual point to estimated match. Results for normalized and unnormalized are shown.
House - unnormalized Residual: 3.0342
House - normalized Residual: 0.21027
Library - unnormalized Residual: 0.3385
Library - normalized Residual: 0.18358
Camera 1 is shown as a blue dot. Camera 2 is a magenta dot. Views are presented from the top and the side of the scene.
House Residual 1: 0.038537 Residual 2: 0.30005
Library Residual 1: 0.16677 Residual 2: 0.26305
Approximately planer points from RANSAC. This is done by using RANSAC to select a dominant inlier subset, then removing them from the match points and repeating.
First dominant set
Second dominant set
Third dominant set
All sets at once