Comparative Analysis of Data Augmentation Techniques in Hand Gesture Recognition

Abstract

Hand gesture recognition has gained significant attention due to its widespread applications in human-computer interaction, virtual reality, and assistive technologies. However, the scarcity of large, labeled datasets poses challenges such as overfitting and limited model generalization. To address this, we systematically evaluate 13 classical and six state-of-the-art (SOTA) data augmentation techniques for hand gesture recognition, conducting experiments on HGR1, OUHANDS, LIBRASUEFS, and EgoHands using the HGR-Net CNN architecture. Our results show that contrast-based augmentations (e.g., Solarize, Invert) improved accuracy by up to 21.16%, while mixing-based methods (e.g., MixUp, CutMix) often reduced performance, likely due to excessive distortion of gesture structures. Additionally, combining the best-performing augmentations was critical for maximizing accuracy across all datasets. However, these combinations did not always produce additive improvements, underscoring the importance of dataset-specific augmentation strategies for achieving optimal model performance.