search

zyong812 / STIP

Code for CVPR22 paper: Exploring Structure-aware Transformer over Interaction Proposals for Human-Object Interaction Detection.
Other
44 stars 6 forks source link

can't reproduce result. #17

Closed OreoChocolate closed 1 year ago

OreoChocolate commented 1 year ago

Hi, thanks for your remarkable work!

I have been attempting to reproduce the v-coco results in Table 3. But, the STIP variations show about 1 AP lower results when compared to the reported results.

I have included my results here. +HM+TR^SS | 63.99 | 68.6 +HM+TR^SC | 63.88 | 68.17 +HM+TR^{SS+SC} | 65.22 | 70.59

Would it be feasible for you to suggest a seed or provide some advice on how I can reproduce the reported results? Thank you in advance!

zyong812 commented 1 year ago

hi, how about using different learning rates?

OreoChocolate commented 1 year ago

Thanks for replying!

I tried changing the learning rates to train the model and got similar results reported in your paper.