search

HRNet / DEKR

This is an official implementation of our CVPR 2021 paper "Bottom-Up Human Pose Estimation Via Disentangled Keypoint Regression" (https://arxiv.org/abs/2104.02300)
MIT License
438 stars 76 forks source link

Replicate experiments for W48 #19

Open bmartacho opened 3 years ago

bmartacho commented 3 years ago

Thanks for sharing your work.

I was able to replicate your results using the w32 weights for COCO, but when using the w48 model, the model doesn't perform over 40% mAP with the pre-trained weights or trying to re-train it. Do you have other correct weights for it or should I use a different protocol for training?

Thank you

Gengzigang commented 3 years ago

Hi, you can try our well-trained model in the model zoo. https://mailustceducn-my.sharepoint.com/personal/aa397601_mail_ustc_edu_cn/_layouts/15/onedrive.aspx?id=%2Fpersonal%2Faa397601%5Fmail%5Fustc%5Fedu%5Fcn%2FDocuments%2FDEKR%2DDisentangled%2DKeypoint%2DRegression%2Fmodel%2Fpose%5Fcoco

bmartacho commented 3 years ago

I have been using the pre-trained model, but differently than the w32, it results in a lower score Screen Shot 2021-11-02 at 09 45 57

Gengzigang commented 3 years ago

I cloned the code and tested the model recently, and the result is the same as the paper. You can use the default parameters and the original code to test it.