Out-of-Distribution Detection in Hand Gesture Recognition Using Image Augmentation

Hand gesture recognition (HGR) is important for creating easy-to-use interfaces that utilize natural human communication and manipulation techniques. To deploy and use HGR in the real world, it is important to detect out-of-distribution (OOD) data. In this study, we perform OOD detection in HGR and show that data augmentation can be a better training technique. Specifically, we utilize hand landmarks coordinates from Mediapipe to add images and expose outliers by flipping the images during the model training process. Our methodology improves FPR95, AUROC performance on NUS hand posture dataset-II. We demonstrate that data augmentation can enable better OOD detection in HGR.