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
207 stars 16 forks source link

Interference #2

Closed Lonicer closed 1 year ago

Lonicer commented 1 year ago

Hello, can I run the code in the run_on_video you provided on my own video like Moment-DETR? If possible, can you provide the best checkpoint of your training and validation process? Thanks!

wjun0830 commented 1 year ago

Although we haven't tried it ourselves, you may be able to.

In addition, since we lost track of our best checkpoints, we provide one available now. There would not be a significant difference. Checkpoints are available here videoonly and video+audio.

Lonicer commented 1 year ago

Although we haven't tried it ourselves, you may be able to.

In addition, since we lost track of our best checkpoints, we provide one available now. There would not be a significant difference. Checkpoints are available here videoonly and video+audio.

Thanks!

Lonicer commented 1 year ago

Excuse me, why are both your code and Moment-DETR saliency prediction scores negative, while the Ground Truth annotations are non-negative? What is the purpose of this setting? Why not add an activation function? Thanks!

勾陈 @.***

 

------------------ 原始邮件 ------------------ 发件人: "wjun0830/QD-DETR" @.>; 发送时间: 2023年4月7日(星期五) 晚上8:31 @.>; @.**@.>; 主题: Re: [wjun0830/QD-DETR] Interference (Issue #2)

Although we haven't tried it ourselves, you may be able to.

In addition, since we lost track of our best checkpoints, we provide one available now. There would not be a significant difference. Checkpoints are available here videoonly and video+audio.

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

wjun0830 commented 1 year ago

I don't think there will be much difference since I think the activation function will only serves to adjust the weight of the loss in that loss function.

Lonicer commented 1 year ago

Hello, sorry to bother you again, how did you confirm the evaluation indicators in the tvsum dataset? I ran your code and found that each category was predicted 5 times. Did you take the average of these five results? Or take the largest of them as the result? Thanks!