Skip to content

Latest commit

 

History

History
180 lines (134 loc) · 5.14 KB

Installation.md

File metadata and controls

180 lines (134 loc) · 5.14 KB
Raw

Installation

Environment

1. Get code

$ git clone https://github.com/lixiny/ArtiBoost.git
$ cd ArtiBoost

2. Set up new environment:

$ conda env create -f environment.yml
$ conda activate artiboost

3. Install dependencies

# inside your artiboost env
$ pip install -r requirements.txt

4. Install thirdparty

  • dex-ycb-toolkit

    $ cd thirdparty
$ git clone --recursive https://github.com/NVlabs/dex-ycb-toolkit.git

    We need install dex-ycb-toolkit as a python package. Following the steps:

    1. you need to install:

      $ sudo apt-get install liboctomap-dev
$ sudo apt-get install libfcl-dev

# or delete the `python-fcl` in dex-ycb-toolkit/setup.py

    2. create a __init__.py in dex_ycb_toolkit

      $ cd thirdparty/dex-ycb-toolkit/dex_ycb_toolkit/
$ touch __init__.py

    3. change a line in dex-ycb-toolkit/setup.py:

      line #16:  opencv-python ==> opencv-python-headless

    finally, at the directory: ./thirdparty, use pip install

    # inside your artiboost env
$ pip install ./dex-ycb-toolkit

    to verify:

    $ python -c "from dex_ycb_toolkit.dex_ycb import DexYCBDataset, _YCB_CLASSES"

Datasets

HO3D

Download HO3D v2 and v3 from the official site. Then unzip and link the datasets in ./data.
Now your ./data folder should have structure like:

    ├── HO3D
    │   ├── evaluation
    │   ├── evaluation.txt
    │   ├── train
    │   └── train.txt
    ├── HO3D_v3
    │   ├── calibration
    │   ├── evaluation
    │   ├── evaluation.txt
    │   ├── manual_annotations
    │   ├── train
    │   └── train.txt

DexYCB

Download DexYCB dataset from the official site. Then unzip and link the dataset in ./data.
Your ./data folder should have structure like:

    ...
    ├── DexYCB
    │   ├── 20200709-subject-01
    │   ├── 20200813-subject-02
    │   ├── 20200820-subject-03
    │   ├── 20200903-subject-04
    │   ├── 20200908-subject-05
    │   ├── 20200918-subject-06
    │   ├── 20200928-subject-07
    │   ├── 20201002-subject-08
    │   ├── 20201015-subject-09
    │   ├── 20201022-subject-10
    │   ├── bop
    │   ├── calibration
    │   └── models

YCB Object Models

Download our pre-processed YCB objects from:

then unzip and copy them to your ./data.

HTML Hand Texture Model

Download our pre-process hand .obj with textures from:

(optional) Download HTML hand texture model from the official site.
then unzip and copy them into ./data.

Finally, you will have ./data with structure like:

    ├── DexYCB
    ├── HO3D
    ├── HO3D_v3
    ├── HTML_release
    │   ├── HTML__hello_world.py
    │   └── ...
    ├── HTML_supp
    │   ├── html_001
    │   ├── ...
    │   ├── html.obj
    │   └── html.obj.mtl
    ├── YCB_models_process
    │   ├── 002_master_chef_can
    │   └── ...
    └── YCB_models_supp
        ├── 002_master_chef_can
        └── ...

Data Assets

Data assets are essential for ArtiBoost training and evaluation.
Download the assets folder at 🔗 here and place it as ./assets.

The ./assets folder should contains:

  • GrabNet/: GrabNet model's weights.
    It is a copy of GrabNet model files/weights from GRAB [Taheri etal ECCV2020]

  • hasson20_assets/:
    This folder contains essentials to run our honetMANO on FPHAB dataset.
    It is a copy of assets folder in handobjectconsist [Hasson etal CVPR2020].

  • postprocess/:
    IKNet model's weights. Convert hand joints position to MANO rotations.
    This checkpoints is trained in the original HandTailor [Lv etal BMVC2021]

  • mano_v1_2/: MANO hand model.
    Download Models & Code at MANO website. Then unzip the downloaded file: mano_v1_2.zip.

  • ho3d_corners.pkl: HO3D object corner's annotation.

  • extend_models_info.json: YCB objects' principal axis of inertia.
    For evaluating maximum symmetry-aware surface distance (MSSD).