[Work Log] Refactoring, cleanup, bug fixes

August 09, 2013
Project Tulips
Subproject Data Association v2
Working path projects/​tulips/​trunk/​src/​matlab/​data_association_2
SVN Revision 15169
Unless otherwise noted, all filesystem paths are relative to the "Working path" named above.


Git Blog Mess

Got sidetracked after screwing up a git commit of the research blog. Not sure the cause, but several files were deleted from the "source" branch. Changed the "preview" rake target so it builds to /tmp/research_blog_site, instead of the source directory. Hopefully this will avoid these issues in the future.

End-to-end experiment

Created an experiment file that recreates yesterday's results from scratch: exp_2013_08_09_animated_reconstruction.m. Since it runs training, it takes about 5 minutes to run.

Also broke out reconstruction code into function in reconstruction/reconstruct_views.m.

Overlay reconstruction onto images

See test/tmp_vis_overlay.m.

I'm seeing some weirdness in the reconstructions.

Stiff Curves Solved

Figured out what was causing stiff curves. I forgot that during training, all precisions are stored with noise_variance fixed at 1.0, and then are scaled on the fly. However, during visualization, that scaling doesn't occur; the precisions are assumed to be stored at the desired scale. i.e. a dumb bug.

The new reconstructions now show a moderate amount of curvature, compared to their pre-bug stiff counterparts.

Found another bug: when constructing the "unscaled" precisions in tr_prep_likelihood.m, I called corr_to_likelihood with params.noise_variance instead of 1.0. Thus, if I understand correctly, the reported training value for noise-variance is 100x lower than it should be.

That means the no-perturb model has a noise standard deviation on the order of 3.4 pixels and the perturb models are around 2.7. This is closer to the range I was expecting, but I was hoping the perturb model stddev would be closer to 0.5, because it should arise only from pixel rasterization. However, other sources of noise could be the curve detector, and also the cubic spline model might not be expressive enough to capture the model variance.

I think I need to re-run end-to-end training and reconstruction to make sure there aren't any side-effects of these fixes.

Committed to revision 15169


Posted by Kyle Simek
blog comments powered by Disqus