search

sunshinglight / DAML_PoseEstimation

MIT License
1 stars 0 forks source link

Cannot repeat the results on SURREAL->LSP #1

Open dkrivosic opened 4 days ago

dkrivosic commented 4 days ago

Hello,

I ran the training for domain adaptation from SURREAL to LSP twice. I got average PCK 79.6 and 79.7 respectively which is way below 84.4 reported in the paper.

Performance of the best checkpoint is worse on each group of keypoints comapred to the model from the paper: Head: 0.052 (-) Shoulders: 67.5 (78.2) Elbows: 82.5 (86.6) Wrists: 80.2 (83.7) Hips: 82.5 (87.1) Knees: 83.3 (85.2) Ankles: 82.5 (85.5)

I ran the code as provided in this repository. Is there something that needs to be changed to get the reported results?

The link to pretrained style transfer models is broken so there could be difference between the models you used and the models I found online. Could you share the style transfer model weights you used?

Thanks

sunshinglight commented 4 days ago

Hi， I found the latest style pretrained-model url in here Did you change the trade-off parameters in your training?

dkrivosic commented 3 days ago

I used the default trade-off parameters, and ran the training script using the command from the readme file. Training the model with the weights you shared didn't improve performance.