brown-ivl / manus

[CVPR 2024] MANUS: Markerless Grasp Capture using Articulated 3D Gaussians
https://ivl.cs.brown.edu/research/manus.html
14 stars 5 forks source link

How to get MANO shape and pose parameters from meta_data? #9

Open ZhaoLizz opened 1 month ago

ZhaoLizz commented 1 month ago

Hi! Thanks for sharing the great work! I have noticed that we can get meta_data, which contains the hand pose in terms of Bones, from the .h5 files. And there are MANO meshes in the evals folder. May I ask how to get MANO pose and shape parameters from meta_data for other frames not in eval? Thanks in advance.

coreqode commented 1 month ago

Hi, MANO parameters are present for grasps sequences in the mano_params directory inside each grasp sequence dir.

ZhaoLizz commented 1 month ago

Thank you for your timely reply.

I have found the mano_params for the grasps sequences. However, I could not locate the mano_params for the hand images in the actions_hdf5, which I intend to use to train a hand model without an object. It appears that the hand images in actions_hdf5 are segmented from the grasps with SAM and InstantNGP.

Is there a correspondence index file available to obtain the mano_params for these hand images in hdf5 from the grasps sequences? This would be extremely helpful for my research.

Thank you very much for your assistance.

coreqode commented 1 month ago

actions_hdf5 is not a part of grasps data. Instead, we captured it individually for each subject to fit hand model.

ZhaoLizz commented 1 month ago

Well, in that case, is there any way I can achieve mano_params for this individual hand-capture data? It seems you have already annotated 3D key points for these hand images but do not provide fitted mano.

ZhaoLizz commented 1 month ago

By the way, can you provide any hint on how to use mano_params from the grasps sequence to generate posed MANO meshes? The mano_params provide angle (33, ), which differs from the original MANO pose parameters (16, 3).

image
coreqode commented 1 month ago

We optimize for PCA components directly. So there are PCA components. You can still get the MANO mesh using this. I will try to update the direction to use MANO and other helper files to use dataset by this weekend.

coreqode commented 14 hours ago

Sorry for delay in response. Please check scripts/dataset_helpers/load_videos.py file on using MANO parameters.