wjun0830 / CGDETR

Official pytorch repository for CG-DETR "Correlation-guided Query-Dependency Calibration in Video Representation Learning for Temporal Grounding"
https://arxiv.org/abs/2311.08835
Other
118 stars 11 forks source link

poor performance on TACoS dataset #21

Open Psycho-9 opened 2 weeks ago

Psycho-9 commented 2 weeks ago

I reproduced your work locally on Video Moment Retrieval dataset, I used the charades-sta and tacos datasets provided by your homepage link, the result on charades-sta is normal, but the result on tacos is abnormal ,the loss didn't converge and the R1iou0.5 is almost 0 , is this a problem about TACoS dataset?

wjun0830 commented 2 weeks ago

We havent faced such an issue. Have you set clip len appropriately?

Psycho-9 commented 2 weeks ago

How to choose a appropriate clip length according to dataset feature ? Now my cli len is 2, is that appropriate?

wjun0830 commented 2 weeks ago

Yes 2 is correct for Tacos. You can compare the feature length and the video length (sec). Which codebase are you using? QD? CG?

Psycho-9 commented 1 week ago

Actually I'm using UVCOM which uses QD as basecode, but the author didn't reply to me.

wjun0830 commented 6 days ago

Hmm.. Im sorry that I havent faced such an issue

awkrail commented 1 day ago

@Psycho-9 Hi, if you have any troubles, could you try our repo? I've reproduced TaCoS experiments w/ CG-DETR.