Open Palm Hand Images Dataset - 500,000 Images for Hand Detection and Gesture Recognition

Overview
The Open Palm Hand Images Dataset is a large-scale commercial biometric dataset produced by Unidata, designed to support hand detection, palm recognition, and gesture analysis research. It includes 500,000 annotated images collected from 50,000 participants, making it one of the largest datasets of its kind for computer vision model training.
Each subject contributes a standardized set of 10 files:
- 6 open-palm photos (live hand images)
- 2 printed-hand images
- 2 replay videos

The dataset covers both right hands and left hands in various configurations, enabling comprehensive research into hand tracking, pose estimation, and gesture recognition across a wide range of real-world conditions.

Subject Demographics
The dataset represents broad demographic diversity across 50,000 participants. Both male and female subjects are included. Each record is labeled with the following metadata attributes:
1. Age
2. Gender
3. Ethnicity
4. Profession (or previous job)
5. Device type used for capture
6. Dominant hand
7. Jewelry presence (has_jewelry)

This level of granularity makes the dataset well-suited for building unbiased computer vision models that generalize across diverse human populations.

Data Collection & Technical Specifications
Data was collected via crowdsourcing platforms under consistent capture protocols, ensuring natural variation in hand appearance, skin tone, and presentation style while maintaining reliable image quality.

1. File format: JPG
2. Labeling type: Metadata per subject (7 attributes)
3. Total records: 500,000 images from 50,000 people
4. Files per subject: 10 (6 live palm photos, 2 printed-hand, 2 replay)

The dataset is fully compliant with GDPR and global data protection regulations. This is a real-world dataset — no synthetic hands or generated data.

Annotation & Labeling
Every sample in the dataset is paired with structured metadata supporting detailed model training and analysis:

1. Age — enables age-stratified model evaluation
2. Gender — male/female split for balanced training data
3. Ethnicity — supports ethnically diverse recognition algorithms
4. Dominant hand — distinguishes left-hand and right-hand subjects
5. Jewelry status — accounts for visual variance from rings, bracelets, and other accessories
6. Profession — additional demographic context
7. Device type — captures hardware diversity across collection devices

Use Cases
- Biometric Authentication and Palm Recognition. The dataset provides real-world human palm images for training palm recognition systems used in identity verification and access control. The three-type structure — live photos, printed images, and replay samples — directly mirrors real attack scenarios, making this dataset especially valuable for building robust anti-spoofing pipelines. Models can learn to distinguish genuine palms from presentation attacks before deployment in security systems.
- Presentation Attack Detection (PAD) A key strength of this dataset is its built-in attack coverage. Beyond real palm captures, each subject contributes:
1. Print attacks — a photo of the palm printed on paper
2. Replay attacks — a photo of the palm displayed on an electronic device
This three-class structure (real / print / replay) enables training and benchmarking of liveness detection models directly on palm data, supporting anti-spoofing research without additional data collection.

Dataset Composition Summary
- 500,000 total annotated images
- 50,000 unique participants
- 10 files per subject (6 palm photos + 2 printed + 2 replay)
- 7 metadata attributes per record
- Both right hands and left hands represented
- Real-world captures — no synthetic data