What's the KLD loss used for?

[sunshineatnoon]

Hi! Thanks for open-sourcing this repo. I found there's a KLD loss (https://github.com/hirokiyokoyama/tracking_via_colorizing/blob/master/train.py#L108), this seems not mentioned by the paper, so could you shed some lights on the function of this loss? Thanks!

[hirokiyokoyama]

I'm not sure if it's still a problem, but, yes, this part is what I added. I wanted to weight the training samples so that ones with large difference in color distribution between reference and target frames are ignored. In such situations (e.g. in a disco, where lighting frequently switches), the copy mechanism cannot work. Conversely, in an ideal situation, colors are only shuffled between frames, so their distributions should be similar.