search

wjun0830 / QD-DETR

Official pytorch repository for "QD-DETR : Query-Dependent Video Representation for Moment Retrieval and Highlight Detection" (CVPR 2023 Paper)
https://arxiv.org/abs/2303.13874
Other
183 stars 13 forks source link

About saliency score #41

Closed zay1817 closed 1 month ago

zay1817 commented 1 month ago

May I ask why the result of my training ”pred_saliency_scores“ is always negative?

wjun0830 commented 1 month ago

We guess that is because the rank contrastive iteratively forces most of the samples to have lower values. We thought that having negative values is not much a problem since highlight detection aims to capture the frames with relatively higher saliency scores. So normalizing the per instance value is an easy way to convert negatives into positive values.

zay1817 commented 1 month ago

Thanks for your reply.