[Work Log] Crash recovery; Index estimation - the saga continues
Machine crashed; need to get matlab environment back to where it was yesterday
TODO:
- load save tracks files
- re-run training (compare against any recorded values)
- re-run wacv with hessian disabled
Done. Process saved in setups/setup_workspace_2014_04.m for future reference.
Index estimation
Using wacv-2012/test_optim.m to visualize the correspondence between triangulated points and reconstructed curve. This is a rough proxy for visualizing curve indices -- bad indices result in a bad reconstruction, and the matching will look bizarre.
...
After much debugging, found that test_optim.m had a regressed, and was displaying a bad reconstruction
I learned an excellent approach for visually assessing indices: run test_optim with centered_mode=true and split_views=true. The resulting plot will connect triangulated points to posterior points, WITH an arrow indicating the backprojection direction. For each curve, rotate the plot so the backprojection arrow disappears (looking in the backprojection direction) -- the connecting line for that point should be nearly perpendicular to the reconstructed curve if the reconstruction is good.
A second key observation: although minimizing the marginal likelihood does result in index shrinkage, the ratio of distances between adjacent points is equal to the ratio of differences between adjacent indices. Thus, the scale component
Third: When bad index estimates cause terrible recponstructions, the per-view reconstructions actually look not bad (i.e. not marginalizing over offset perturbations). Position of per-view reconstructed curves is suspect.
Posted by Kyle Simek