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
724 stars 127 forks source link

Mano trans parameter #94

Closed Bozcomlekci closed 3 years ago

Bozcomlekci commented 3 years ago

Greetings,

I couldn't understand why trans parameter is passed to the mano layer for extracting joints and vertices. Some implementations (e. g. MeshTransforner - METRO) were not passing this parameter. What is the function of this parameter?

Also, could you explain why root wrist is subtracted from mano_joint_cam but not from mano_mesh_cam?

Thanks in advance.

mks0601 commented 3 years ago

trans is just a 3D translation vector. You can add the 3D translation vector to the output 3D joint/mesh coordinates if you do not add them. mano_mesh_cam needs some translation for rendering.

Bozcomlekci commented 3 years ago

Thanks for the explanation for mano_trans.

Maybe I should have asked it in a separate issue but I still have questions regarding mano_joint_cam and mano_mesh_cam. Why don't you perform these operations on the mano_mesh_cam as well? In this line, you switch back to milimeter. Doesn't it create an inconsistency between mano_mesh_cam and mano_joint_cam since you apply these operations only one of them?

mks0601 commented 3 years ago

As you might notice I do not pass mano_mesh_cam to targets in here.

Bozcomlekci commented 3 years ago

I got it. Thanks for the reply.