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
290 stars 39 forks source link

what is the difference for data with and without "occ"? #28

Open Wuchuq opened 2 years ago

Wuchuq commented 2 years ago

Thank you for your work! For the pre-processed data with name "occ" and without, what's the difference of them?

hongsukchoi commented 2 years ago

As written in the implementation detail (Section 4 of the paper), we followed "How robust is 3d human pose estimation to occlusion?".

It is occluding images with PASCAL objects in a random way.

QitaoZhao commented 2 years ago

I have a similar question. What is the difference between the data with "tight" and without for Human3.6m. Thanks for answering.

Wuchuq commented 2 years ago

thank you for your response!

hongsukchoi commented 2 years ago

@QitaoZhao

It’s the tightness of bounding boxes used for training. We preprocessed another version to compare with SPIN setting that used tight bboxes