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
817 stars 147 forks source link

Wrong camera translation params of human36m #58

Closed KaifuZ closed 4 years ago

KaifuZ commented 4 years ago

The camera translation in all Human36M_subject_camera.json seem to be wrong. for example: in Human36M_subject1_camera.json, 't' of '1' is [-346.05078140028075, 546.9807793144001, 5474.481087434061], but in matlab_meta.txt of s_01_act_02_subact_01_ca_01, the translation is 1841.107028 4955.284623 1563.445396.

mks0601 commented 4 years ago

Maybe the original matlab_meta.txt file has t = -RT. I modified t of json file to follow new world2cam function .

The old one follows world_cood = R(cam_coord-t). The new one follows world_coord - R*cam_coord+ t.

I modified the world2cam function and all camera parameters in the google drive. If you are using old json file, you should use the old world2cam function. If you are using the new one, you should use the new function.

KaifuZ commented 4 years ago

thx!