Closed wanboyang closed 2 years ago
Yes. It is for multi-label training videos used for center loss.
“if (labels[i] > 0).sum() == 0 or ((labels[i] > 0).sum() != 1 and itr < itr_th): continue” In my opinion, It means the center-loss only work on single-label training video before the itr up to the itr_th(itr_th=5000 in this paper). Can you notice that what happened in the experiment if itr_th=0?
Yes, we introduce the multi-label videos after a while, so that the learning of the centers is easy for training in the beginning. Generally, center loss in image classification tasks is introduced after learning the classifier to a reasonable extent. That is why we introduce the multi-label video after some iters, than at the beginning. Though I don't remember the exact values for Itr_th=0, it did have a performance drop.
the
if (labels[i] > 0).sum() == 0 or ((labels[i] > 0).sum() != 1 and itr < itr_th): continue
inDoes it mean center-loss use on multi-label training video?