Hi,
I am trying to reproduce your results on TAD using ActionFormer on THUMOS, I downloaded your features from huggingface following this link and just replaced the I3D features used in ActionFormer with yours, and obtained the following results:
│ tIoU │ mAP │
| ---- | ---- |
│ 0.3 │ 84.07 |
│ 0.4 │ 79.45 |
│ 0.5 │ 71.96 |
│ 0.6 │ 62.06 |
│ 0.7 │ 46.09 |
│ avg │ 68.73 |
Your report 69.6 in your repo here (https://github.com/OpenGVLab/VideoMAEv2/blob/29eab1e8a588d1b3ec0cdec7b03a86cca491b74b/docs/TAD.md)
I was wondering if there's anything you tweaked from the THUMOS I3D-based training original recipe ? for reference https://github.com/happyharrycn/actionformer_release/blob/main/configs/thumos_i3d.yaml
I noticed there were some changes w.r.t the original recipe in the InternVideo1 repo (https://github.com/OpenGVLab/InternVideo/blob/main/InternVideo1/Downstream/Temporal-Action-Localization/configs/thumos.yaml), namely the number of epochs moving from 30 (ActionFormer) to 45 (InternVideo1).
Did you also change the number of epochs for VideoMAEv2?
Hi,
I am trying to reproduce your results on TAD using ActionFormer on THUMOS, I downloaded your features from huggingface following this link and just replaced the I3D features used in ActionFormer with yours, and obtained the following results:
Your report 69.6 in your repo here (https://github.com/OpenGVLab/VideoMAEv2/blob/29eab1e8a588d1b3ec0cdec7b03a86cca491b74b/docs/TAD.md)
I was wondering if there's anything you tweaked from the THUMOS I3D-based training original recipe ? for reference https://github.com/happyharrycn/actionformer_release/blob/main/configs/thumos_i3d.yaml
I noticed there were some changes w.r.t the original recipe in the InternVideo1 repo (https://github.com/OpenGVLab/InternVideo/blob/main/InternVideo1/Downstream/Temporal-Action-Localization/configs/thumos.yaml), namely the number of epochs moving from 30 (ActionFormer) to 45 (InternVideo1).
Did you also change the number of epochs for VideoMAEv2?