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

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

The provided freihand pose param is different from freihand origin dataset

EAST-J opened this issue · comments

Hi, I download the origin mano params from the freihand. I found that the provided dataset have the different pose_param, but they have the same shape_params. I wonder what's the difference between them?

That is related with 'flat_hand_mean' option of MANO. https://github.com/hassony2/manopth/blob/4f1dcad1201ff1bfca6e065a85f0e3456e1aa32b/manopth/manolayer.py#L34

Yeah, I notice lmb-freiburg/freihand#6 (comment) this issue, I set the flat_hand_mean to False and visualize the mesh, both origin pose_params and provided pose_params seem ok to me.
![orign](https://user-images.githubusercontent.com/51024280/172368374-1e48145a-f062-463f-9241-dd7b80544ed8.PN
provided
But I compare the pose_param between the origin and provided, there is a small deviation for the first three params(I subtract these two terms)
tmp
I wonder how this deviation come about or how can I get the provided pose_params from the original one?

The first three params represent 3D root rotation (3D wrist rotation). The original FreiHAND dataset does not provide 3D root rotation in the camera-centered coordinate system. Instead, it provides 3D root rotation in the world-coordinate system. I obtained the 3D root rotation in the camera centered coordinate system by rigidly aligning 3D hand meshes to the provided 3D joint coordinates in the camera-centered coordinate system. That is why there is a small 3D root rotation difference.