"A Meta-Learning Framework for Few-Shot Dynamic Hand Gesture Recognition with Soft Temporal-aware Contrastive Learning"

"This paper presents a comprehensive dataset designed to evaluate the robustness of hand gesture recognition (HGR) against dynamic upper-limb postures. The dataset includes multi-channel surface Electromyography (sEMG) and concurrent tri-axial Accelerometer (ACC) signals, acquired from 10 subjects. Data was collected using an 8-channel wireless Trigno system, sampling sEMG at 1259 Hz and ACC at 148 Hz. The sensing configuration comprised six electrodes placed equidistantly around the mid-forearm and two placed on the extensor and flexor carpi radialis.The experimental protocol requires subjects to perform 11 static hand gestures (e.g., Fist, Pinch, Wrist Flexion) while executing 12 distinct dynamic upper-limb postures. These postures simulate daily activities, such as raising the arm, pouring, drinking, and target-reaching movements.To systematically evaluate model generalization, the data is structured across four sessions collected on different days:Session 1: A baseline dataset (7 gestures, 8 postures, 3 repetitions) for model pre-training.Session 2: Data collected after a controlled 0.5 cm electrode shift, using an expanded set of 10 functional gestures (8 postures, 3 repetitions) to test robustness to spatial and temporal variations.Session 3: A single-repetition dataset (10 gestures, 8 postures) intended for few-shot fine-tuning and testing.Session 4: A single-repetition dataset (10 gestures, 4 new dynamic postures) to evaluate model generalization to unseen movements.We provide raw signals .This dataset facilitates research in robust HGR, particularly for developing and benchmarking algorithms against common challenges like electrode shift, inter-day variations, and novel limb postures, with specific utility for pre-training and few-shot learning strategies."