search

HowieMa / PPT

[ECCV 2022] "PPT: token-Pruned Pose Transformer for monocular and multi-view human pose estimation"
55 stars 1 forks source link

High MPJPE value on s9 subject after 20 epochs. #4

Open Louis-w1214 opened 1 year ago

Louis-w1214 commented 1 year ago

Thank you for you work! I have used your code for training 20 epochs and found bad performance (high MPJPE value, over 150mm) on s9 subject but good performance on s11 subject, so what could be the reason?

HowieMa commented 1 year ago

Hi, Thanks for your interest in our work! I uploaded our trained model to Google Drive, here is the link. You can also see the evaluation results log inside the folder. I would suggest you evaluate this model first. If you still meet the same issue, there may be some issues with your dataset, as it is a little strange to have a good performance on one subject but a terrible performance on the other.

HowieMa commented 1 year ago

Another possibility is that there are some issues with S9, as mentioned in https://arxiv.org/pdf/1905.05754.pdf image. Thus, I am not sure if you preprocess these data correctly.

We follow the default setting in Epipolar Transformer and remove these damaged images as well, you can see this code for more detail.

Louis-w1214 commented 1 year ago

Thank you for your commond, I have fixed this problem! Additionly, I noticed that in the multi-view pretrained coco model(ppt_s_ratio_07_coco_256_256.pth), the resolution is 256x256 which is different form the resolution in single-view model. I wonder what you have done to resize the figure from 256x192 resolution to 256x256. I would appreciate it if you can release the code of pretraining the model on the 256x256 coco dataset.