[Work Log] WACV results
WACV deadline has passed, but still working on reconstructing ground truth, because it's a good test on the end-to-end system.
Open issues:
- weird issue with branching
- clear up math issue with branching
Branching bug
Fixed! It was a bug in infer_branch_points() function. I was assuming Corr.curve_sm used the same index set as Corr.ll_means_flat, which it was not.
I wanted to use Corr.curve_sm, because it is precomputed in corr_to_likelihood(), and it would save me an extra semi-expensive call to the cubic spline smoother, csaps(). In the end, I modified corr_to_likelihood to save the smoothing amount and used it to call csaps inside run_wacv() to re-smooth the curves. At least csaps is faster (linear time) than computing the posterior mean (quadratic).
Test Results
Results on dataset 2:
Looks good!
All Results
Running on all 11 datasets...
Matlab is running out of memory on the second dataset (apparently)
When running directly on second dataset, it's fine. Is the pass on the first dataset leaking memory or something?
Maybe one of the mex files?
It's weird, because we've run on dataset #2 tens of times without this problem...
Now running directly on dataset #2 is resulting in a curve of over 1 million points. investigating...
Solution: the image dimensions were reversed (should be nrows X ncols, not width x height). The triangulations were nutso, and after projecting and resampling, the number of points was in the millions. It's interesting that it's so easy to make this algorithm blow up; are there other; will need to keep an eye out for more benign ways to trigger this kind of explosion.
Ground truth issues
Encountered this error message when running on all ground truth datasets:
Warning: file wacv-2012/datasets/3/ground_truth_2d.gt2
Missing curves in view 6: 1, 2, 3, 4, 5, 6, 7, 8, 10, 11, 12, 13, 14, 15, 16,
Warning: file wacv-2012/datasets/6/ground_truth_2d.gt2
Missing curves in view 2: 2, 3,
Warning: file wacv-2012/datasets/6/ground_truth_2d.gt2
Missing curves in view 3: 3,
Warning: file wacv-2012/datasets/6/ground_truth_2d.gt2
Missing curves in view 4: 2, 3,
Warning: file wacv-2012/datasets/6/ground_truth_2d.gt2
Missing curves in view 6: 2, 3,
Warning: file wacv-2012/datasets/6/ground_truth_2d.gt2
Missing curves in view 8: 1, 3,
Warning: file wacv-2012/datasets/6/ground_truth_2d.gt2
Missing curves in view 10: 3,
Warning: file wacv-2012/datasets/6/ground_truth_2d.gt2
Missing curves in view 15: 3,
Warning: file wacv-2012/datasets/6/ground_truth_2d.gt2
Missing curves in view 19: 3,
Warning: file wacv-2012/datasets/6/ground_truth_2d.gt2
Missing curves in view 23: 2, 3,
Warning: file wacv-2012/datasets/6/ground_truth_2d.gt2
Missing curves in view 27: 3,
Warning: file wacv-2012/datasets/6/ground_truth_2d.gt2
Missing curves in view 31: 3,
Warning: file wacv-2012/datasets/6/ground_truth_2d.gt2
Missing curves in view 35: 3,
Dataset 6 looks to be totally crap; need to investigate.
Found a bug in corr_to_likelihood that led to out-of-bound indexing.
Looks like Liu never finished ground-truthing dataset 2010-01-26/arab_1_36.
Skipping for now; lets see if others are okay...
Dataset 9 is broken. investigating...
Dataset 9 seems not terrible, actually. A few missing views on the hard-to-trace views of a few curves.
Some cuves were so short that they contained only one point, which violated an assert that requires all curves to be at least 2 points long. Adjusted the bezier-to-sampled-curve code so that the final point is always included.
Finished running all. issues
- dataset 1 is very rough
- datasets 6 was skipped
- datasets 7 - 11 are worthless.
Need to re-trace 6 through 11.
Finished re-trace of 6. Whew, 90 minutes... that sucked. Reconstruction looks good now.
TODO
- Finish tracing datasets 7 through 11.
- Work on attached-curve-multi-view prior and reconstruction.