search

WuJie1010 / Temporally-language-grounding

A Pytorch implemention for some state-of-the-art models for" Temporally Language Grounding in Untrimmed Videos"
96 stars 21 forks source link

revise iou to softmax(iou) #5

Closed yufansong closed 5 years ago

yufansong commented 5 years ago

I try to add a softmax operation to predict IoU. In the paper, it doesn't do this operation, though, during the training process, I find after I fix the bug about

Predict_IoUs = torch.zeros(opt.num_steps, batch_size)
Predict_IoUs[step, :] = tIoU
iou_loss += torch.abs(Previous_IoUs[i,j] - Predict_IoUs[i,j])

The predict iou will always not in 0-1. That will make the network get a really really bad performance, even training a long time with the pre-trained model. For IoU problem, if choose to use a softmax operation, that will be not so bad.

Actually, several weeks ago, when I fix the "predict iou" and "location" bug, the performance will be worse. It is a very strange thing and the author of the paper didn't reply to my email.