The following datasets are used for training.
We have used webdataset to pack the datasets into shards for efficient training. In training POEM, we have also filtered samples with right hand only. The packed dataset tars used for training and evaluation is released here. After downloading the dataset, create a directory data
in the root directory and organize the files as follows:
├── assets
├── config
├── data
│ ├── dataset_tars
│ ├── Oakink_mv
│ ├── HO3D_mv
│ ├── HO3D_mv_test
│ ├── ...
│ └── Arctic_mv
├── exp
├── lib
├── ...
Download mano_v1_2.zip
from the MANO website (Sign in -> Download -> Models & Code), unzip, and copy it to assets/mano_v1_2
:
$ cd assets
$ cp -r path/to/mano_v1_2 assets/