Some improvements to the approach outlined at the end of previous post...
Tweak epipolar likelihood to add a small affine and nonlinear distortion component (e.g. due to lens distortion). this allows epipolar variance to be smaller; forces deviations to be spatially correlated.
Find some good correspondences; fix camera distortion approach 1: distort image using GP and known correspondences. approach 2: optimize camera distortion parameters under GP and known correspondences
Use unambiguous correspondences first.
I fear that graph topology may be too noisy and fluxuating to be a reliable source of information. Is keypoint matching with epipolar constraint enough to solve this? Maybe with added spatial distortion?
How can adding correspondences reduce uncertainty in remaining points? The approach above can reduce uncertainty to almost lying on the epipolar line. If we can recover small subgraphs, those can help too. But will false joints break things? (this would be a good experiment) Maybe just linear sections with interior keypoint nodes. This should at least improve fine-scale matching later.
2D gives us lateral branch candidates, but are ambiguous. Can resolve them in 3D?
Idea: branch points have higher geodesic distance?
Simpler problem: Camera repair for volumetric reconstruction.
Sift keypoint matching nonlinear distortion fix ransac matching keypoints + distortion correction
Posted by Kyle Simek