BiCE-HG (A bi-conditional egocentric hand gesture dataset under controlled environmental conditions)

BiCE-HG (A bi-conditional egocentric hand gesture dataset under controlled environmental conditions) is a first-person view (FPV) RGB hand gesture dataset collected for augmented reality (AR) and virtual reality (VR) interaction research. The dataset captures 19 distinct hand gestures — 15 single-hand and 4 bimanual — performed by 23 participants across six environmental conditions formed by the factorial combination of two lighting levels (full lighting and low lighting) and three motion scenarios (sitting, standing, and walking).Data were recorded at 3840 × 2160 resolution at 30 FPS using a head-mounted camera providing an egocentric perspective. Each gesture recording provides: raw RGB frames, MediaPipe-annotated frames with hand skeleton overlay, per-frame 3D hand keypoints (21 landmarks × up to 2 hands, normalised x/y/z coordinates), a single representative static frame with corresponding keypoints, and raw and annotated video clipsHand keypoints were extracted using a multi-stage MediaPipe Hands pipeline with additional image preprocessing to improve robustness under challenging lighting and motion conditions. The dataset contains 123,175 frames across 2,607 recordings, achieving an overall frame-level detection rate of 98.44%. Detection rates across the six conditions range from 95.69% (low lighting, standing) to 99.41% (full lighting, walking), demonstrating robustness across all experimental conditions.The 19 gesture classes are organised into 12 action categories covering touch and selection (pinch, air tap), hand shapes (open palm, fist front, fist back), navigation (horizontal swipe, vertical swipe), rotation (wrist rotate), symbolic gestures (thumbs up, OK sign), pointing, greeting (wave, V sign), drawing (clockwise and counter-clockwise circles), zoom (zoom in, zoom out), and framing (hand frame left up, hand frame right up).The dataset is provided in two versions: a dynamic version (keypoint sequences for temporal gesture recognition) and a static version (one representative frame per gesture per recording for pose-based classification). Both versions include 3D hand keypoint files in plain-text format and are accompanied by Python data loaders compatible with PyTorch. An 80/20 train/test split stratified by gesture is recommended for reproducible evaluation.All participants provided written informed consent. Ethics approval was granted by the University of Sheffield, School of Electrical and Electronic Engineering (Reference: 067003, approved 08/07/2025)