ddshan/hint

HInt dataset from HaMeR: Reconstructing Hands in 3D with Transformers

HInt: Hand Interactions in the wild

Data repository for the paper: Reconstructing Hands in 3D with Transformers

Georgios Pavlakos, Dandan Shan, Ilija Radosavovic, Angjoo Kanazawa, David Fouhey, Jitendra Malik

Overview

The HInt dataset is contributed in the paper, Reconstructing Hands in 3D with Transformers, with the goal to complement existing datasets used for training and evaluation 3D hand pose estimation.

HInt annotates 2D keypoint locations and occlusion labels for 21 keypoints on the hand. It is built off of 3 existing datasets (Hands23, Epic-Kitchens VISOR, and Ego4D) and provides annotations for images from the three existing datasets.

Prepare HInt Dataset

Step 1: download partial HInt

Download Hint_annotation_partial.zip which contains all annotations, New Days and Epic-Kitchens VISOR image frames. That's said, this zip file contains everything except the Ego4D frames duo to Ego4D license constraint.

wget https://fouheylab.eecs.umich.edu/~dandans/projects/hamer/HInt_annotation_partial.zip
unzip HInt_annotation_partial.zip

After unzip, the folder structure will be as below. Each folder contains the .jpg image and .json annotation pairs except Ego4D directories (noted with * at the end) are missing .jpg frames. We provide instructions on how to retrieve Ego4D frames in Step 2 down below.

HInt_annotation_partial
├── TEST_ego4d_img*
├── TEST_ego4d_seq*
├── TEST_epick_img
├── TEST_newdays_img
├── TRAIN_ego4d_img*
├── TRAIN_epick_img
├── TRAIN_newdays_img
├── VAL_ego4d_img*
├── VAL_ego4d_seq*
├── VAL_epick_img
└── VAL_newdays_img

Step 2: prepare Ego4D frames

1. Get access. Follow the Start Here page on Ego4D official website to get download access.

The process will be like: submit your information form -> wait for the mail about the agreement -> review and accept the terms of Ego4D license agreement. If your license agreement is approved, you will receive an email from Ego4D about the AWS access credentials. As it mentioned, this process might take ~48 hours so do it earlier.

2. Set up Ego4D CLI. Follow Ego4D Dataset Download CLI to set up your CLI to get ready for downloading.

3. Download Ego4D clips. The clips will be saved under /path/to/ego4d_data/v1/clips

ego4d --output_directory="/path/to/save/ego4d_data" --version v1 --datasets clips annotations --metadata 

4. Decode Ego4D clips. Set the ego4d_root and hint_root in the argparse first. The decoded clips will be saved under /path/to/ego4d_data/v1/clips_decode.
   This script is dependent on ffmpeg library, you can install it by conda install ffmpeg=5.1.2 (a tested version).

cd prep_HInt
python prep_ego4d.py --task=decode_clips

5. Retrieve Ego4D frames. Fill Ego4D frames in Ego4D folders under HInt_annotation_partial. Once it passed the file amount checking, the dataset folder name HInt_annotation_partial will be updated to HInt_annotation.

python prep_ego4d.py --task=retrieve_frames

6. Check MD5 to verify data integrity. Compare MD5 of zip files between your generated HInt and the original one. Make sure you pass it first especially before you use the Ego4D subset of HInt.

python prep_ego4d.py --task=verify_hint

Visualize HInt Annotations

Plot annotations on images. This script is dependent on mmengine library, you can install it simply by pip install mmengine.

cd visualize_HInt
python draw_hand.py

Citing

If you find this data useful for your research, please consider citing the following paper. If you have questions about the dataset, feel free to email Dandan Shan.

HaMeR

@inproceedings{pavlakos2024reconstructing,
    title={Reconstructing Hands in 3{D} with Transformers},
    author={Pavlakos, Georgios and Shan, Dandan and Radosavovic, Ilija and Kanazawa, Angjoo and Fouhey, David and Malik, Jitendra},
    booktitle={CVPR},
    year={2024}
}

Epic-Kitchens VISOR

@article{darkhalil2022epic,
  title={Epic-kitchens visor benchmark: Video segmentations and object relations},
  author={Darkhalil, Ahmad and Shan, Dandan and Zhu, Bin and Ma, Jian and Kar, Amlan and Higgins, Richard and Fidler, Sanja and Fouhey, David and Damen, Dima},
  journal={Advances in Neural Information Processing Systems},
  volume={35},
  pages={13745--13758},
  year={2022}
}

Ego4D

@inproceedings{grauman2022ego4d,
  title={Ego4d: Around the world in 3,000 hours of egocentric video},
  author={Grauman, Kristen and Westbury, Andrew and Byrne, Eugene and Chavis, Zachary and Furnari, Antonino and Girdhar, Rohit and Hamburger, Jackson and Jiang, Hao and Liu, Miao and Liu, Xingyu and others},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={18995--19012},
  year={2022}
}

hint's Issues

Pre-processing HInt like HaMeR

Thanks for releasing the HInt dataset! I was wondering if you had any scripts to format HInt in the webdataset format for training with HaMeR.

Thanks!
Vimal

The approximate size of HInt dataset

Can you please inform me of the size of the HInt dataset?

Ego4D download

Hi, thanks for sharing this dataset. I try to download Ego4D data with

ego4d --output_directory="ego4d_data" --version v1 --datasets clips annotations --metadata 

This command will download 739G data as follows

Is there any way to avoid downloading files of such a large size?