search

vchoutas / smplify-x

Expressive Body Capture: 3D Hands, Face, and Body from a Single Image
https://smpl-x.is.tue.mpg.de/
Other
1.73k stars 336 forks source link

it seems when use the model only on upper body , the results are terrible #54

Closed visonpon closed 4 years ago

visonpon commented 4 years ago

Even the keypoints from openpose are natural, the output mesh still terrible, anybody can explain why? thanks~

geopavlakos commented 4 years ago

For initialization, SMPLify-X uses the two shoulder joints and the two hip joints. If the hip joints are not detected (as I suspect that is the case in the upper body examples you are looking at), then the translation and the global rotation are estimated using only two joints, which is an underconstrained problem and it is probably the reason for the weird results you are reporting. With that being said, even if the hip joints are detected, detecting the knees and/or ankles is useful to put more constraints on the body pose and return more reasonable results.