[Work Log] Work Log

April 07, 2014

I've been experimenting with alternative methods for inferring the index set and have several good options. When equipped with a good index set, it seems most of the problems we've been seeing disappear. As sad as it seems, its seeming like index optimization should be tossed out, despite all the time I spent deriving the equations.

I've observed that too high of a perturb smoothing variance causes the mean curve to drift toward the linear-initial model. Manually lowering perturb smoothing variance and raising smoothing variance causes better reconstruction and much higher marginal likelihood.

Centering matters -- too low of a position_variance causes rate variance to take over, which causes weird results in some cases.

Should refact tr_train to use new index estimation scheme; goal: getting better higher smoothness variance and lower perturb_smoothness variance

Open question: is camera linearization beneficial in training?

Refactored proces_tracks to recieve index estimation method.

Refactored run_wacv_5 to use iterative index estimation.

Posted by Kyle Simek
blog comments powered by Disqus