search

Arthur151 / ROMP

Monocular, One-stage, Regression of Multiple 3D People and their 3D positions & trajectories in camera & global coordinates. ROMP[ICCV21], BEV[CVPR22], TRACE[CVPR2023]
https://www.yusun.work/
Apache License 2.0
1.36k stars 231 forks source link

3DPW results #471

Closed fabienbaradel closed 1 year ago

fabienbaradel commented 1 year ago

Hi @Arthur151 , Thanks for your amazing work. In the supp. mat. of the BEV paper you are reporting result of ROMP and BEV on 3DPW test set (Table 5). But since your method is solving at the same time the detection and the regression I am wondering what kind of detection rate do you get? I am not able to find this information in your paper. And for Table 5 (cf attached) do you compute the metrics on the matched person or on all the persons? Thanks for you help,

Screenshot 2023-07-25 at 14 52 26

Arthur151 commented 1 year ago

@fabienbaradel During evaluation, we directly match the ground truth with the most close prediction for evaluation. So I think this refers to all people. As we all know that 3DPW only provide the ground truth for 1 or 2 people in the scene. In many scenes, it lacks detection annotations for all people in the scene, while ROMP and BEV try to predict all people in the scene. So, you know.

Thanks for your interests in our work! We really appreciate it. Best, Yu