[Work Log] Work Log

December 27, 2013


New energy function causes much worse index ordering. Ordering before optimization is terrible; new energy function seems to introduce new local minima that we can't overcome.

Running old version to recall problems.


Significant offset from LL reconstruction. Increasing perturb_scale improves significantly (from 2.5 to 81). When influsence of wide-baseline views is small, pulling toward camera is significant; increasing perturb_scale increasing that influence.

Side effect of increasing perturb_scale is massive increase of tip extension in the direction of the camera. This is because the posterior is more peaked, decreasing the size of the global optimum.

Adding index smoothness metaprior seems to help, but I'm not confident in the implementation, since there's no unit test for it. Also getting a small amount of index extension, possibly because end-caps arent a full unit away from previous point. can this be fixed? Can't really use chord-length in 2D, because this could contain significant noise. Although using ordinal isn't much different. Could store fraction of last point. Also need to acknowledge that different sub-sampling periods cause GP parameters to change meaning. Should re-number ordinals based on resampling. Store in data?


dataset 6: overextension in teal top curve, red mid curve dataset 7: overextension in dk green top curve dataset 8: extreme overextension in teal top curve; bad ll localization dataset 9: overextension in red top curve

and others. Always when ll point is in direction of previous curve.

Posted by Kyle Simek
blog comments powered by Disqus