search

FraLuca / STSGCN

Repository for "Space-Time-Separable Graph Convolutional Network for Pose Forecasting" (ICCV 2021)
MIT License
94 stars 25 forks source link

Evaluation metrics #13

Closed AndiDemon closed 2 years ago

AndiDemon commented 2 years ago

Hello, It is an awesome work and achievement. Thank you for providing the code. I would like to ask about the unit in the evaluation method. In the paper it is use the mm unit. If I am not wrong by reading the code of the utils/loss_funcs.py, It does not show any unit in mm or pixels. Am I wrong? or do you use the pixels first and then convert it to mm in the paper? Thank you. Best regards.

FraLuca commented 2 years ago

Hi AndiDemon,

We're happy you enjoyed our work! You are right, we use mm unit for the MPJPE loss. In particular, in the loss function, we input skeleton joint positions in Euclidean coordinates (x,y,z). So we never use images or conversion pixel to mm. You can consider the resulting value of the loss in mm. I hope this will clear your doubt. Luca

AndiDemon commented 2 years ago

Thank you for your quick respond, I understand, considering the input which is quite complicated I tried to run the code on CPU yesterday. I need to do some changes on the code regarding the cuda() function. But the program run well after that.

Best regards,

Dean-UQ commented 2 years ago

Dear readers of STSGCN, It is clear that there are some mistakes represented in this paper. Since the author did not declare and revise their published paper until now, it is necessary to reflect this issue to the program chair of ICCV-21. So, please leave your comments here to make the chair know the impartial peer reviews from the human motion prediction community. Thanks for your contributions to the community.