[Work Log] Thoughts on a New approach
- Improve calibration: distortion, bundle adjustment
- MRF-based per-view depth estimates
- merging or killing structure by Maximum marginal likelihood
0. feature extraction hacks
- Use ELSD for segments
- use background subtraction to ignore spurious curves
1. Improve calibration
There's no reason not to model lens distortion and run bundle adjustment. This will allow some form of multi-view stereo to become feasible.
bundle adjustment will let us correct for errors in calibration target (finally!)
2. Initial dept estimate for all curve points
It should be relatively easy to create a simple MRF for depth estimation. Likelihood: nearest reprojected distance. prior: first or second order markove in 1-dimension. Look into MRF solvers (is kolmogorov's code available? TRW?).
alternative: keypoint matching and gap-filling. think pmvs, but with 2 dof instead of 3
alternative:
3. merging, branching, and pruning using maximum marginal likleihood
greedily pick pairs to merge, using DTW to match points and branching gaussian process to perform marginal likleihood. compare against mcmc
future papers could add branching and background rejection.
4. reconstruct
maximum posterior
5. evaluate
compare against hand-picked correspondences
compare against bundler + pmvs.
other ideas
**Alternarive initial depth estimate: sift keypoint matching for seeding multi-view stereo. use background model to mask-out background keypoints.
ceres-solver for nonlinear least squares task.
evaluate sensitivity to poor calibration. which methods deteriorate fastest?
Posted by Kyle Simek