search

hongsukchoi / TCMR_RELEASE

Official Pytorch implementation of "Beyond Static Features for Temporally Consistent 3D Human Pose and Shape from a Video", CVPR 2021
MIT License
287 stars 39 forks source link

How to get smpl parameters annotation in Human3.6M processing? #38

Open Wuchuq opened 1 year ago

Wuchuq commented 1 year ago

Thank you for your work! But how to get the smpl annotation commented here in the preprocessing of human3.6M dataset image

hongsukchoi commented 1 year ago

@Wuchuq

I uploaded the file with smpl param. So you can just uncomment the lines. If you want the smpl param data before parsing, you can obtain them from the neural annot homepage.

Wuchuq commented 1 year ago

Thank you so much!

Wuchuq commented 1 year ago

I find that the smpl annotation is incomplete for all the frames. For act_01_subact_01, there are only 277 annotations available. But in the preprocessed human3.6M data you provided, the smpl parameters are available for all the frames. So is there any problems with the smpl_param file? image

hongsukchoi commented 1 year ago

Hi,

I couldn't understand the question. Do you mean the smpl annotation from the NeuralAnnot homepage is incomplete?

Wuchuq commented 1 year ago

Yes, seems like the annotation is only for the frames after 5 times downsampling.

hongsukchoi commented 1 year ago

Got it. Normally image-based 3DHPSE methods use subsampled data, since the images are not much different in near frames. Human3.6M images are in 50Hz. So we released subsampled annotation for them in the NeuralAnnot homepage.

Wuchuq commented 1 year ago

Got it. Normally image-based 3DHPSE methods use subsampled data, since the images are not much different in near frames. Human3.6M images are in 50Hz. So we released subsampled annotation for them in the NeuralAnnot homepage.

Got it! Thanks.