oakink / tink

The official implementation of Tink -- one of the core contributions in the CVPR2022 paper: OakInk.

Tink

CVPR, 2022
Lixin Yang* · Kailin Li* · Xinyu Zhan* · Fei Wu · Anran Xu . Liu Liu · Cewu Lu
\star = equal contribution

This repo contains the official implementation of Tink -- one of the core contributions in the CVPR2022 paper: OakInk.

Tink is a novel method that Transfers the hand's INteraction Knowledge among objects.

Installation

• First, clone this repo:

  git clone https://github.com/KailinLi/Tink.git
  cd Tink
  git submodule init && git submodule update

• Second, to set up the environment, follow the instruction: stand-alone in OakInk to install the environment with conda.

• Third, inside the OakInk directory, install the oikit as package:

  $ cd OakInk
  $ pip install .

Download

In this repo, we provide a mini dataset to demonstrate the pipeline of Tink.

• Download the assets files.
• Download mano following the official instructions. And put the mano_v1_2 under the assets directory.
• Download the mini dataset from this link. And unzip them under the DeepSDF_OakInk directory.

Your directory should look like this:

Tink
├── assets
│   ├── anchor
│   ├── hand_palm_full.txt
│   └── mano_v1_2
├── DeepSDF_OakInk
│   ├── data
│   │   ├── meta
│   │   ├── OakInkObjects
│   │   ├── OakInkVirtualObjects
│   │   ├── raw_grasp
│   │   └── sdf
│   │       └── phone

DeepSDF

In this section, we demonstrate how to preprocess the object meshes and train a category-level DeepSDF.

If you are not interested in training DeepSDF, feel free to skip this section.

1. Compile the C++ code

Please follow the official instructions of DeepSDF.

You will get two executables in the DeepSDF_OakInk/bin directory. (We modified some of the original source code in DeepSDF, so please make sure to compile these scripts from the scratch.)

2. Preprocess the object meshes

export MESA_GL_VERSION_OVERRIDE=3.3
export PANGOLIN_WINDOW_URI=headless://

cd DeepSDF_OakInk
python preprocess_data.py --data_dir data/sdf/phone --threads 4

After finishing the script, you can find the SDF files in DeepSDF_OakInk/data/sdf/phone/SdfSamples directory.

3. Train the network

CUDA_VISIBLE_DEVICES=0 python train_deep_sdf.py -e data/sdf/phone

4. Dump the latent codes and reconstructed meshes

CUDA_VISIBLE_DEVICES=0 python reconstruct_train.py -e data/sdf/phone  --mesh_include

You can find the reconstructed meshes under the DeepSDF_OakInk/data/sdf/phone/Reconstructions/Meshes.

Shape Interpolation

If you skip the above section, we provide a pre-trained DeepSDF network. Please download the files, unzip them and replace the original phone directory:

sdf
├── phone
│   ├── network
│   │   ├── ModelParameters
│   │   │   └── latest.pth
│   │   └── LatentCodes
│   ├── Reconstructions
│   │   ├── Codes
│   │   │   ├── C52001.pth
│   │   │   ├── ...
│   │   └── Meshes
│   │       ├── C52001.ply
│   │       ├── ...
│   ├── rescale.pkl
│   ├── SdfSamples
│   │   ├── C52001.npz
│   │   ├── ...
│   ├── SdfSamples_resize
│   ├── specs.json
│   └── split.json

Now, go to the Tink directory, and run the following script to generate the interpolations:

cd ..

# you can generate all of the interpolations:
python tink/gen_interpolate.py --all -d ./DeepSDF_OakInk/data/sdf/phone

# or just interpolate between two objects (from C52001 to o52105):
python tink/gen_interpolate.py -d ./DeepSDF_OakInk/data/sdf/phone -s C52001 -t o52105

You can find the interpolations in DeepSDF_OakInk/data/sdf/phone/interpolate directory.

Calculate Contact Info

We calculate the contact region of C52001:

python tink/cal_contact_info.py \
	-d ./DeepSDF_OakInk/data/sdf/phone \
	-s C52001 \
	--tag demo \
	-p DeepSDF_OakInk/data/raw_grasp/demo/C52001_0001_0000/2021-10-09-15-13-39/dom.pkl \
	--vis

The contact_info.pkl is stored in DeepSDF_OakInk/data/sdf/phone/contact/C52001/demo_e54965ec08. e54965ec08 is the hash code of the hand parameters.

Contact Mapping

We take the virtual object o52105 as an example.

To transfer the contact information from C52001 to o52105:

python tink/info_transform.py \
	-d ./DeepSDF_OakInk/data/sdf/phone \
	-s C52001 \
	-t o52105 \
	-p DeepSDF_OakInk/data/sdf/phone/contact/C52001/demo_e54965ec08

You can find the transfered contact info in DeepSDF_OakInk/data/sdf/phone/contact/C52001/demo_e54965ec08/o52105.

Pose Refinement

CUDA_VISIBLE_DEVICES=0 python tink/pose_refine.py \
	-d ./DeepSDF_OakInk/data/sdf/phone \
	-s C52001 \
	-t o52105 \
	-p DeepSDF_OakInk/data/sdf/phone/contact/C52001/demo_e54965ec08 \
	--vis

The fitted hand pose will be stored in DeepSDF_OakInk/data/sdf/phone/contact/C52001/demo_e54965ec08/o52105 directory. (When visualizing the hand pose, you might need to chick the 'x' on the window status bar to start fitting.)

We also provide all the transferred hand poses of the mini dataset. You can download the files, unzip them and replace the original phone directory.