search

mks0601 / I2L-MeshNet_RELEASE

Official PyTorch implementation of "I2L-MeshNet: Image-to-Lixel Prediction Network for Accurate 3D Human Pose and Mesh Estimation from a Single RGB Image", ECCV 2020
MIT License
708 stars 130 forks source link

Freihand annotation differs from the original anotation #96

Closed biansy000 closed 2 years ago

biansy000 commented 2 years ago

Hello, after I download the freihand annotation you provided here, I find that the annotation is almost the same with the original dataset I downloaded from freihand, except that the first three items of theta (refers to global orientation) are different. Is it possible for you to explain why the difference exists? Thanks if you could help.

mks0601 commented 2 years ago

Hi,

In the original annotation, I remember that the global rotation is a rotation to world coordinate system. I changed that to a rotation to camera coordinate system.

biansy000 commented 2 years ago

Could you explain how the rotation is derived? As I do not see any available information in the original dataset I downloaded from freihand.

mks0601 commented 2 years ago

They provide 3D hand joint coordinates in camera-centered coordinate system. I rigidly align the mesh to the 3D joint coordinates.

biansy000 commented 2 years ago

I understand the process now, thanks for your quick reply!