Learning to Disambiguate Strongly Interacting Hands via Probabilistic Per-Pixel Part Segmentation [3DV 2021 Oral]

Learning to Disambiguate Strongly Interacting Hands via Probabilistic Per-Pixel Part Segmentation,
Zicong Fan, Adrian Spurr, Muhammed Kocabas, Siyu Tang, Michael J. Black, Otmar Hilliges International Conference on 3D Vision (3DV), 2021

Features

DIGIT estimates the 3D poses of two interacting hands from a single RGB image. This repo provides the training, evaluation, and demo code for the project in PyTorch Lightning.

Updates

November 25 2021: Initial repo with training and evaluation on PyTorch Lightning 0.9.

Setting up environment

DIGIT has been implemented and tested on Ubuntu 18.04 with python >= 3.7, PyTorch Lightning 0.9 and PyTorch 1.6.

Clone the repo:

git clone https://github.com/zc-alexfan/digit-interacting

Create folders needed:

make folders

Install conda environment:

conda create -n digit python=3.7
conda deactivate
conda activate digit
conda install pytorch==1.6.0 torchvision==0.7.0 cudatoolkit=10.1 -c pytorch
pip install -r requirements.txt

Downloading InterHand2.6M

Download the 5fps.v1 of InterHand2.6M, following the instructions here
Place annotations, images, and rootnet_output from InterHand2.6M under ./data/InterHand/*:

./data/InterHand
├── annotations
├── images
│   ├── test
│   ├── train
│   └── val
├── rootnet_output
│   ├── rootnet_interhand2.6m_output_all_test.json
│   └── rootnet_interhand2.6m_output_machine_annot_val.json
|-- annotations
|-- images
|   |-- test
|   |-- train
|   `-- val
`-- rootnet_output
    |-- rootnet_interhand2.6m_output_test.json
    `-- rootnet_interhand2.6m_output_val.json

The folder ./data/InterHand/annotations should look like this:

./data/InterHand/annotations
|-- skeleton.txt
|-- subject.txt
|-- test
|   |-- InterHand2.6M_test_MANO_NeuralAnnot.json
|   |-- InterHand2.6M_test_camera.json
|   |-- InterHand2.6M_test_data.json
|   `-- InterHand2.6M_test_joint_3d.json
|-- train
|   |-- InterHand2.6M_train_MANO_NeuralAnnot.json
|   |-- InterHand2.6M_train_camera.json
|   |-- InterHand2.6M_train_data.json
|   `-- InterHand2.6M_train_joint_3d.json
`-- val
    |-- InterHand2.6M_val_MANO_NeuralAnnot.json
    |-- InterHand2.6M_val_camera.json
    |-- InterHand2.6M_val_data.json
    `-- InterHand2.6M_val_joint_3d.json

Preparing data and backbone for training

Download the ImageNet-pretrained backbone from here and place it under:

./saved_models/pytorch/imagenet/hrnet_w32-36af842e.pt

Package images into lmdb:

cd scripts
python package_images_lmdb.py

Preprocess annotation:

python preprocess_annot.py

Render part segmentation masks:

Following the README.md of render_mano_ih to prepare an LMDB of part segmentation. For question in preparing the segmentation masks, please keep issues in there.

Place the LMDB from the images, the segmentation masks, and meta_dict_*.pkl to ./data/InterHand and it should look like the structure below. The cache files meta_dict_*.pkl are by-products of the step above.

|-- annotations
|   |-- skeleton.txt
|   |-- subject.txt
|   |-- test
|   |   |-- InterHand2.6M_test_MANO_NeuralAnnot.json
|   |   |-- InterHand2.6M_test_camera.json
|   |   |-- InterHand2.6M_test_data.json
|   |   |-- InterHand2.6M_test_data.pkl
|   |   `-- InterHand2.6M_test_joint_3d.json
|   |-- train
|   |   |-- InterHand2.6M_train_MANO_NeuralAnnot.json
|   |   |-- InterHand2.6M_train_camera.json
|   |   |-- InterHand2.6M_train_data.json
|   |   |-- InterHand2.6M_train_data.pkl
|   |   `-- InterHand2.6M_train_joint_3d.json
|   `-- val
|       |-- InterHand2.6M_val_MANO_NeuralAnnot.json
|       |-- InterHand2.6M_val_camera.json
|       |-- InterHand2.6M_val_data.json
|       |-- InterHand2.6M_val_data.pkl
|       `-- InterHand2.6M_val_joint_3d.json
|-- cache
|   |-- meta_dict_test.pkl
|   |-- meta_dict_train.pkl
|   `-- meta_dict_val.pkl
|-- images
|   |-- test
|   |-- train
|   `-- val
|-- rootnet_output
|   |-- rootnet_interhand2.6m_output_test.json
|   `-- rootnet_interhand2.6m_output_val.json
`-- segm_32.lmdb

Training and evaluating

To train DIGIT, run the command below. The script runs at a batch size of 64 using accumulated gradient where each iteration is on a batch size 32:

python train.py --iter_batch 32 --batch_size 64 --gpu_ids 0 --trainsplit train --precision 16 --eval_every_epoch 2 --lr_dec_epoch 40 --max_epoch 50 --min_epoch 50

OR if you just want to do a sanity check you can run:

python train.py --iter_batch 32 --batch_size 64 --gpu_ids 0 --trainsplit minitrain --valsplit minival --precision 16 --eval_every_epoch 1 --max_epoch 50 --min_epoch 50

Each time you run train.py, it will create a new experiment under logs and each experiment is assigned a key.

Supposed your experiment key is 2e8c5136b, you can evaluate the last epoch of the model on the test set by:

python test.py --eval_on minitest --load_ckpt logs/2e8c5136b/model_dump/last.ckpt

python test.py --eval_on test --load_ckpt logs/2e8c5136b/model_dump/last.ckpt

The former only does the evaluation 1000 images for a sanity check.

Similarly, you can evaluate on the validation set:

python test.py --eval_on val --load_ckpt logs/2e8c5136b/model_dump/last.ckpt

Visualizing and evaluating pre-trained DIGIT

Here we provide instructions to show qualitative results of DIGIT.

Download pre-trained DIGIT:

wget https://dataset.ait.ethz.ch/downloads/dE6qPPePCV/db7cba8c1.pt
mv db7cba8c1.pt saved_models

Visualize results:

CUDA_VISIBLE_DEVICES=0 python demo.py --eval_on minival --load_from saved_models/db7cba8c1.pt  --num_workers 0

Evaluate pre-trained digit:

CUDA_VISIBLE_DEVICES=0 python test.py --eval_on test --load_from saved_models/db7cba8c1.pt --precision 16
CUDA_VISIBLE_DEVICES=0 python test.py --eval_on val --load_from saved_models/db7cba8c1.pt --precision 16

You should have the same results as in here.

The results will be dumped to ./visualization.

Citation

@inProceedings{fan2021digit,
  title={Learning to Disambiguate Strongly Interacting Hands via Probabilistic Per-pixel Part Segmentation},
  author={Fan, Zicong and Spurr, Adrian and Kocabas, Muhammed and Tang, Siyu and Black, Michael and Hilliges, Otmar},
  booktitle={International Conference on 3D Vision (3DV)},
  year={2021}
}

License

Since our code is developed based on InterHand2.6M, which is CC-BY-NC 4.0 licensed, the same LICENSE is applied to DIGIT.

DIGIT is CC-BY-NC 4.0 licensed, as found in the LICENSE file.

References

Some code in our repo uses snippets of the following repo:

Our initial codebase was based on InterHand2.6M and we made significant modification to it.
We use UNet code from UNet-PyTorch
We use HRNet backbone from HRNet-Human-Pose-Estimation

Please consider citing them if you find our code useful:

@inproceedings{Moon_2020_ECCV_InterHand2.6M,  
author = {Moon, Gyeongsik and Yu, Shoou-I and Wen, He and Shiratori, Takaaki and Lee, Kyoung Mu},  
title = {InterHand2.6M: A Dataset and Baseline for 3D Interacting Hand Pose Estimation from a Single RGB Image},  
booktitle = {European Conference on Computer Vision (ECCV)},  
year = {2020}  
}  

@inproceedings{sun2019deep,
  title={Deep High-Resolution Representation Learning for Human Pose Estimation},
  author={Sun, Ke and Xiao, Bin and Liu, Dong and Wang, Jingdong},
  booktitle={CVPR},
  year={2019}
}

@inproceedings{xiao2018simple,
    author={Xiao, Bin and Wu, Haiping and Wei, Yichen},
    title={Simple Baselines for Human Pose Estimation and Tracking},
    booktitle = {European Conference on Computer Vision (ECCV)},
    year = {2018}
}

@misc{Charles2013,
  author = {milesial},
  title = {Pytorch-UNet},
  year = {2021},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/milesial/Pytorch-UNet}}
}

Contact

For any question, you can contact zicong.fan@inf.ethz.ch.

human2b/digit-interacting