A Smartphone Camera Based RGB Video Dataset of Natural Hand Gestures
Description
This dataset provides RGB video recordings and corresponding frame sequences for 13 everyday hand gestures, captured using a Samsung Galaxy S23 smartphone under real-world lighting conditions. Designed for research on gesture recognition using only consumer-grade cameras, the dataset excludes depth or infrared sensors to emphasize feasibility with standard hardware. Each gesture was performed by two individuals, across three takes, under both bright daylight and dim green LED lighting. To capture different spatial contexts, recordings were made in two views: hand-only (focused solely on the hand) and full-body (showing the entire participant). This results in a total of 312 standardized video clips: 13 gestures × 2 participants × 3 takes × 2 lighting conditions × 2 capture modes. Each video was processed to a uniform resolution of 640×640 pixels at 30 fps, and frames were extracted into organized image sequences. The data is arranged by gesture class and capture mode, with a labels.csv file at the root providing metadata for each clip. This includes gesture label, participant ID, lighting condition, view type, and take number, along with the corresponding frame path. The dataset is ideal for training and evaluating machine learning models for gesture recognition, particularly in scenarios that require robustness to lighting variations, user diversity, and mobile deployment.
Files
Steps to reproduce
1. Download all the zipped gesture class folders and the `labels.csv` file from the dataset repository. 2. Unzip each gesture class folder into a single root directory. After extraction, the directory structure should contain one folder per gesture class (e.g., 'come_here/', 'fist/', etc.). 3. Place the provided `labels.csv` file in the same root directory (alongside the gesture class folders). 4. Each gesture class folder contains two subfolders: - `hand_only/` - `full_body/` Each of these contains: - `videos/` (.mp4 clips) - `frames/` (folders with extracted .png images) 5. The `labels.csv` file provides metadata and mappings for each video, including: - Gesture label - Participant ID - Lighting condition - Capture mode (hand_only or full_body) - Take number - Corresponding frame path 6. No additional processing is needed. Users can directly read `labels.csv` and use it to link video files with corresponding frame sequences and annotations. 7. No special software is required beyond a standard unzipping tool. To programmatically interact with the data, users may optionally use Python with libraries such as `pandas` and `os`.
Institutions
- American International University Bangladesh