search

happyharrycn / actionformer_release

Code release for ActionFormer (ECCV 2022)
MIT License
419 stars 77 forks source link

Why are the reported results of TadTR on Thumos14 and ActivityNet quite different with those in TadTR paper? #62

Closed ttgeng233 closed 1 year ago

ttgeng233 commented 1 year ago

Thank you for your good job! I found that the results are very different between your paper and TadTR paper on Thumos14 and ActivityNet datasets? Are there some specific implementation details?

bcac8b0e67166b446d72a53b9516ca4 ac8282c74c2b5aad330719c2e9e2c96

happyharrycn commented 1 year ago

These results are from the arXiv version of TadTR during the submission of the paper. This older version can be accessed from this link on arXiv. TadTR has subsequently updated the results (with the latest arXiv version this Aug), which we were not able to include in the ECCV camera ready. We are happy to incorporate their latest results in our arXiv version (probably after all conference deadlines this fall).

happyharrycn commented 1 year ago

I will mark this closed for now unless further questions arise.