HaGRID - HAnd Gesture Recognition Image Dataset

We introduce a large image dataset HaGRIDv2 (HAnd Gesture Recognition Image Dataset) for hand gesture recognition (HGR) systems. You can use it for image classification or image detection tasks. Proposed dataset allows to build HGR systems, which can be used in video conferencing services (Zoom, Skype, Discord, Jazz etc.), home automation systems, the automotive sector, etc.

HaGRIDv2 size is 1.5T and dataset contains 1,086,158 FullHD RGB images divided into 33 classes of gestures and a new separate "no_gesture" class, containing domain-specific natural hand postures. Also, some images have no_gesture class if there is a second gesture-free hand in the frame. This extra class contains 2,164 samples. The data were split into training 76%, 9% validation and testing 15% sets by subject user_id, with 821,458 images for train, 99,200 images for validation and 165,500 for test.

The dataset contains 65,977 unique persons and at least this number of unique scenes. The subjects are people over 18 years old. The dataset was collected mainly indoors with considerable variation in lighting, including artificial and natural light. Besides, the dataset includes images taken in extreme conditions such as facing and backing to a window. Also, the subjects had to show gestures at a distance of 0.5 to 4 meters from the camera.

Example of sample and its annotation:

For more information see our arxiv paper [TBA]().

🔥 Changelog

• 2024/09/24: We release HaGRIDv2. 🙏
  • The HaGRID dataset has been expanded with 15 new gesture classes, including two-handed gestures
  • New class "no_gesture" with domain-specific natural hand postures was addad (2,164 samples, divided by train/val/test containing 1,464, 200, 500 images, respectively)
  • Extra class no_gesture contains 200,390 bounding boxes
  • Added new models for gesture detection, hand detection and full-frame classification
  • Dataset size is 1.5T
  • 1,086,158 FullHD RGB images
  • Train/val/test split: (821,458) 76% / (99,200) 9% / (165,500) 15% by subject user_id
  • 65,977 unique persons
• 2023/09/21: We release HaGRID 2.0. ✌️
  • All files for training and testing are combined into one directory
  • The data was further cleared and new ones were added
  • Multi-gpu training and testing
  • Added new models for detection and full-frame classification
  • Dataset size is 723GB
  • 554,800 FullHD RGB images (cleaned and updated classes, added diversity by race)
  • Extra class no_gesture contains 120,105 samples
  • Train/val/test split: (410,800) 74% / (54,000) 10% / (90,000) 16% by subject user_id
  • 37,583 unique persons
• 2022/06/16: HaGRID (Initial Dataset) 💪
  • Dataset size is 716GB
  • 552,992 FullHD RGB images divided into 18 classes
  • Extra class no_gesture contains 123,589 samples
  • Train/test split: (509,323) 92% / (43,669) 8% by subject user_id
  • 34,730 unique persons from 18 to 65 years old
  • The distance is 0.5 to 4 meters from the camera

Installation

Clone and install required python packages:

git clone https://github.com/hukenovs/hagrid.git
# or mirror link:
cd hagrid
# Create virtual env by conda or venv
conda create -n gestures python=3.11 -y
conda activate gestures
# Install requirements
pip install -r requirements.txt

Downloads

We split the train dataset into 34 archives by gestures because of the large size of data. Download and unzip them from the following links:

Dataset

dataset annotations: annotations

HaGRIDv2 512px - lightweight version of the full dataset with min_side = 512p 119.4 GB

or by using python script

python download.py --save_path <PATH_TO_SAVE> \
                   --annotations \
                   --dataset

Run the following command with key --dataset to download dataset with images. Download annotations for selected stage by --annotations key.

usage: download.py [-h] [-a] [-d] [-t TARGETS [TARGETS ...]] [-p SAVE_PATH]

Download dataset...

optional arguments:
  -h, --help            show this help message and exit
  -a, --annotations     Download annotations
  -d, --dataset         Download dataset
  -t TARGETS [TARGETS ...], --targets TARGETS [TARGETS ...]
                        Target(s) for downloading train set
  -p SAVE_PATH, --save_path SAVE_PATH
                        Save path

After downloading, you can unzip the archive by running the following command:

unzip <PATH_TO_ARCHIVE> -d <PATH_TO_SAVE>

The structure of the dataset is as follows:

├── hagrid_dataset <PATH_TO_DATASET_FOLDER>
│   ├── call
│   │   ├── 00000000.jpg
│   │   ├── 00000001.jpg
│   │   ├── ...
├── hagrid_annotations
│   ├── train <PATH_TO_JSON_TRAIN>
│   │   ├── call.json
│   │   ├── ...
│   ├── val <PATH_TO_JSON_VAL>
│   │   ├── call.json
│   │   ├── ...
│   ├── test <PATH_TO_JSON_TEST>
│   │   ├── call.json
│   │   ├── ...

Models

We provide some models pre-trained on HaGRIDv2 as the baseline with the classic backbone architectures for gesture classification, gesture detection and hand detection.

In addition, if you need to detect hands, you can use YOLO detection models, pre-trained on HaGRIDv2

However, if you need a single gesture, you can use pre-trained full frame classifiers instead of detectors. To use full frame models, remove the no_gesture class

Demo

python demo.py -p <PATH_TO_CONFIG> --landmarks

Demo Full Frame Classifiers

python demo_ff.py -p <PATH_TO_CONFIG>

Annotations

The annotations consist of bounding boxes of hands and gestures in COCO format [top left X position, top left Y position, width, height] with gesture labels. We provide user_id field that will allow you to split the train / val / test dataset yourself, as well as a meta-informations contains automatically annotated age, gender and race.

"04c49801-1101-4b4e-82d0-d4607cd01df0": {
    "bboxes": [
        [0.0694444444, 0.3104166667, 0.2666666667, 0.2640625],
        [0.5993055556, 0.2875, 0.2569444444, 0.2760416667]
    ],
    "labels": [
        "thumb_index2",
        "thumb_index2"
    ],
    "united_bbox": [
        [0.0694444444, 0.2875, 0.7868055556, 0.2869791667]
    ],
    "united_label": [
        "thumb_index2"
    ],
    "user_id": "2fe6a9156ff8ca27fbce8ada318c592b",
    "hand_landmarks": [
            [
                [0.37233507701702123, 0.5935673528948108],
                [0.3997604810145188, 0.5925499847441514],
                ...
            ],
            [
                [0.37388438145820907, 0.47547576284667353],
                [0.39460467775730607, 0.4698847093520443],
                ...
            ]
        ]
    "meta": {
        "age": [24.41],
        "gender": ["female"],
        "race": ["White"]
    }

• Key - image name without extension
• Bboxes - list of normalized bboxes for each hand [top left X pos, top left Y pos, width, height]
• Labels - list of class labels for each hand e.g. like, stop, no_gesture
• United_bbox - united combination of two hand boxes in the case of two-handed gestures ("hand_heart", "hand_heart2", "thumb_index2", "timeout", "holy", "take_picture", "xsign") and 'null' in the case of one-handed gestures
• United_label - a class label for united_bbox in case of two-handed gestures and 'null' in the case of one-handed gestures
• User ID - subject id (useful for split data to train / val subsets).
• Hand_landmarks - auto-annotated with MediaPipe landmarks for each hand.
• Meta - automatically annotated with FairFace and MiVOLO neural networks meta-information contains age, gender and race

Bounding boxes

Landmarks

Converters

License

This work is licensed under a variant of Creative Commons Attribution-ShareAlike 4.0 International License.

Please see the specific license.

Authors and Credits

• Alexander Kapitanov
• Andrey Makhlyarchuk
• Karina Kvanchiani
• Aleksandr Nagaev
• Roman Kraynov
• Anton Nuzhdin

Links

• Github
• arXiv

Citation

You can cite the paper using the following BibTeX entry:

@InProceedings{Kapitanov_2024_WACV,
    author    = {Kapitanov, Alexander and Kvanchiani, Karina and Nagaev, Alexander and Kraynov, Roman and Makhliarchuk, Andrei},
    title     = {HaGRID -- HAnd Gesture Recognition Image Dataset},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month     = {January},
    year      = {2024},
    pages     = {4572-4581}
}