search

DART2022 / DART

DART: Articulated Hand Model with Diverse Accessories and Rich Textures (NeurIPS 2022 - Datasets and Benchmarks Track)
https://dart2022.github.io/
132 stars 9 forks source link

Finger penetration issue in the dataset #8

Closed xmlyqing00 closed 1 year ago

xmlyqing00 commented 1 year ago

Hi,

Thanks for providing such a large hand dataset. May I ask about the pose code format in the released dataset? It seems that they're in the MANO format. which is different from the paper that says you generate hand pose by A-MANO. I used the MANO decoder to decode the hand pose to mesh but found finger penetration issues. And some poses look unnatural.

Thanks for your time!!!

image image image

tomguluson92 commented 1 year ago

Hi @xmlyqing00 , Thanks for you mention this problem. We do some pose adaptation as to fit with the code we use in DARTset.py. Please use that code for any comparison and things like that.

lixiny commented 1 year ago

@xmlyqing00 May I ask how are the parameters configured for your MANO layer?

I suggest that using the ManoLayer from manotorch package with the following parameters:

ManoLayer(
    joint_rot_mode="axisang",
    use_pca=False,
    mano_assets_root="assets/mano_v1_2",
    center_idx=0,
    flat_hand_mean=False
)

and check whether these issues still persist?

xmlyqing00 commented 1 year ago

Hi, I used the wrong decoder to decode the hand code. Now the penetration problem is not an issue. Thanks for your help!

Best