search

mks0601 / 3DMPPE_POSENET_RELEASE

Official PyTorch implementation of "Camera Distance-aware Top-down Approach for 3D Multi-person Pose Estimation from a Single RGB Image", ICCV 2019
MIT License
807 stars 147 forks source link

Questions about evaluation on Human36m dataset？ #93

Open ShirleyMaxx opened 3 years ago

ShirleyMaxx commented 3 years ago

hi, thanks for your great work!

I'm confused about the evaluation on the H36M dataset. If I don't misunderstand, the predicted root-relative pose should add root depth predicted from RootNet to get the final estimated absolute pose, which is then aligned to root to calculate the Protocol 2 MPJPE.

So it is not clear for me in the below code. Why do you add the ground-truth root depth before back projecting to the camera coordinate system? It seems that this evaluation process uses additional GT root information and is not fair when comparing to other methods, like Sun (ECCV'18).

https://github.com/mks0601/3DMPPE_POSENET_RELEASE/blob/3f92ebaef214a0eb1574b7265e836456fbf3508a/data/Human36M/Human36M.py#L174

Looking forward to your reply! Thanks in advance!

mks0601 commented 3 years ago

gt_3d_root is from here, which is the output of RootNet.

ShirleyMaxx commented 3 years ago

oh, got it! There is no problem! Thank you again!