README.md

Harmonious Feature Learning for Interactive Hand-Object Pose Estimation

Directory

${ROOT}  
|-- data  
|   |-- HO3D
|   |   |-- data
|   |   |   |-- train
|   |   |   |   |-- ABF10
|   |   |   |   |-- ......
|   |   |   |-- evaluation
|   |   |   |-- train_segLable
|   |   |   |-- ho3d_train_data.json
|   |-- DEX_YCB
|   |   |-- data
|   |   |   |-- 20200709-subject-01
|   |   |   |-- ......
|   |   |   |-- object_render
|   |   |   |-- dex_ycb_s0_train_data.json
|   |   |   |-- dex_ycb_s0_test_data.json

Data

You need to follow directory structure of the data as below.

• Download HO3D(version 2) data data
• Download DexYCB data data
• Download the process data data

Pytorch MANO layer

• For the MANO layer, I used manopth. The repo is already included in manopth.
• Download MANO_RIGHT.pkl and MANO_LEFT.pkl from here and place at assets/mano_models.

Train

HO3d

sh sh/train_ho3d.sh

Dex-ycb

sh sh/train_dex-ycb.sh

Test

HO3d

sh sh/train_ho3d_test.sh

Dex-ycb

sh sh/train_dex-ycb_test.sh

Acknowledgments

We thank:

• Semi-Hand-Object
• HandOccNet