USTC-Video-Understanding/I3D_Finetune

Reproduce results

giladsharir opened this issue · 7 comments

giladsharir commented

Hi,
I'm trying to reproduce the results reported in the paper (and this repository) with the I3D model trained with only RGB on UCF-101 (split 1) .
When I run test.py with the checkpoint provided in this repository (finetuned on UCF-101 RGB), I get a score of 89.4 . The authors of this repository report 94.7 .
Do you have any idea why is there a discrepancy in the scores?

Catchher commented

Aha，I meet the same issue, particularly in Optical-Flow branch with only 83% acc.

vra commented

Hi @giladsharir and @Catchher , Thanks for trying our code and I am sorry for the inconvenience. I guess we have uploaded wrong checkpoints and we are figuring it out. Thanks for you patience.

Catchher commented

Thanks for ur reply, I'm looking forward to ur good news. @vra

tangshixiang commented

Hi, could you tell me whether you have uploaded the corrected checkpoints or not? Thank you~ @vra

sumeetssaurav commented

Hi,
I'm trying to reproduce the results reported in the paper (and this repository) with the I3D model trained with only RGB on UCF-101 (split 1) .
When I run test.py with the checkpoint provided in this repository (finetuned on UCF-101 RGB), I get a score of 89.4 . The authors of this repository report 94.7 .
Do you have any idea why is there a discrepancy in the scores?

TaoXieSZ commented

I got 91% as the result, it is normal?

lanfeng4659 commented

Hi,
I'm trying to reproduce the results reported in the paper (and this repository) with the I3D model trained with only RGB on UCF-101 (split 1) .
When I run test.py with the checkpoint provided in this repository (finetuned on UCF-101 RGB), I get a score of 89.4 . The authors of this repository report 94.7 .
Do you have any idea why is there a discrepancy in the scores?

have you settled the problem? I get the 91.5% scores using the shared model trained on the UCF101 RGB.