[Work Log] Pre-tails fix

August 10, 2013
Project Tulips
Subproject Data Association v2
Working path projects/​tulips/​trunk/​src/​matlab/​data_association_2
Unless otherwise noted, all filesystem paths are relative to the "Working path" named above.

Pre-tails issue

Bug: Smallest index of reconstructed curves is significantly greater than 0.


This results in long pre-tails, as seen in this image:



Recall that reconstruction occurs by doing (1) rough triangulation, (2) smoothing, then (3) re-triangulating against the smoothed curve. The initial triangulation usually results in a very bad index set, with spacing far larger than it should be, due to poor localization by maximum likelihood. The subsequent smoothing causes the curve to stretch out longer than it should, so when re-triangulation occurs, ends are cut off.


After re-triangulating, re-index so the minimum index is zero.

Change to correspondence/corr_to_likelihood.m:

145  % re-index by subtracting minimum index
146  min_index = min([Corr.ll_indices{:}]);
147  Corr.ll_indices = cellfun(@(x) x - min_index, Corr.ll_indices, 'UniformOutput', false);


This likely had an effect on training results, because marginal prior variance of the initial point was over-estimated, because it's index was over-estimated.


I'm afraid I made a mess of things yesterday when I was addressing the noise_variance issue in training. Need to review the end-to-end systems for training, reconstruction and marginal likelihood evaluation.


Posted by Kyle Simek
blog comments powered by Disqus