search

Michel-liu / GroupPose

[ICCV 2023] The official PyTorch code for Group Pose: A Simple Baseline for End-to-End Multi-person Pose Estimation
Apache License 2.0
79 stars 1 forks source link

A question on loss of `mix_refpoint` #2

Closed JunrQ closed 1 year ago

JunrQ commented 1 year ago

Hello, thank you for your wonderful work, i have a question,

in https://github.com/Michel-liu/GroupPose/blob/9ed9e7ef4e3b247242700691ca1db92e194b5537/models/grouppose/transformer.py#L167C60-L167C95 ,

the mean of keypoint coordinates is concatenated in front, but the keypoints loss is calculated for first 17/14 keypoints, is that a bug?

Michel-liu commented 1 year ago

Thank you for your report ❤️! I have checked the code and will fix this issue soon.

JunrQ commented 1 year ago

Dear author, I trained the resnet coco and the ap is still 0 after 15 epoches, is that normal?

Michel-liu commented 1 year ago

That's a strange result. Can you provide more information, such as the output log and training args you used? BTW, with the 12 epochs training setting, the AP comes to 56.2 and 60.1 at the 10th and 11th epochs.

JunrQ commented 1 year ago

Thank you for your quick reply, previous i trained with 6 gpus and the model failed to converge, but when i retrained with 8 gpus, it seems normal.

rrenle commented 10 months ago

Thank you for your quick reply, previous i trained with 6 gpus and the model failed to converge, but when i retrained with 8 gpus, it seems normal.

Hi, can you reproduce the results of the paper? My result is AP=70.5 which is lower than 72.