A 1.5-terabyte hello to hand gesture recognition

What it does

HaGRIDv2 is a large-scale image dataset for hand gesture recognition. It contains 1,086,158 FullHD RGB images covering 33 gesture classes plus a no_gesture catch-all, collected from 65,977 unique subjects indoors under varied lighting. The maintainers also provide pre-trained YOLO-based detectors and classifiers as baselines.

The interesting bit

The authors trained a dynamic gesture recognition algorithm entirely on these static images, which lives in a separate repository. That is a neat trick—teaching a model to understand motion without ever showing it video.

Key highlights

• 1.5 TB of FullHD RGB imagery, with a lighter 512 px version weighing in at ~119 GB
• 65,977 unique subjects, shot at distances from 0.5 to 4 meters under mixed artificial and natural light
• Splits are stratified by user_id, so the same person never appears across train and test sets
• Pre-trained baseline models included, with a YOLOv10x gesture detector reaching 89.4 mAP
• Annotations include bounding boxes and landmarks

Caveats

• The full dataset is split into 34 separate zip archives, so prepare for a download marathon
• The dynamic gesture recognition code lives in a separate repository, not here
• Subjects are all over 18 and mainly indoors, so generalization to minors or outdoor scenes is untested

Verdict

Worth a look if you are building hand-gesture controls for video conferencing, automotive, or smart-home interfaces and need real-world training data. Skip it if you were hoping for a quick, lightweight download or dynamic gesture code out of the box.