search

facebookresearch / co-tracker

CoTracker is a model for tracking any point (pixel) on a video.
https://co-tracker.github.io/
Other
2.72k stars 195 forks source link

Questions about loss computation #94

Closed qsisi closed 2 months ago

qsisi commented 3 months ago

Hello! @nikitakaraevv

In the balanced CE loss implementations: https://github.com/facebookresearch/co-tracker/blob/9ed05317b794cd177674e681321780614a65e073/cotracker/models/core/cotracker/losses.py#L14-L38 For simplicity, assume a 1-d scenario where gt = [1, 0], pred=[p0, p1]∈(0,1), then as computed above: pos = [1, 0] neg = [0, 1] label = [1, -1] a = [-p0, p1] b = [0, p1] loss1 = b + log(exp(-b) + exp(a-b)) = [log(1+exp(-p0)), p1 + log(1+exp(-p1))] My question is, why not just use this cross-entropy computation:image then the loss will be like: loss2 = [-log(p0), -log(1-p1)] ? What is the difference between these two loss computations? Is the first one better than the second one?

qsisi commented 3 months ago

Maybe it's because the first one is smoother than the second one? image

qsisi commented 2 months ago

let pred==gt, the balanced_ce_loss is not equal to 0: image which can be explained by the above curves, that the computed loss will not equal zero.

nikitakaraevv commented 2 months ago

Hi @qsisi, that's a great question, thank you! Yes, this loss is indeed always greater than zero. We've run some experiments and found out that the model converges better with this loss, but we're trying to fix this for the next version.