Implicit Semantic Data Augmentation for Hand Pose Estimation

Data augmentation is a well-known technique used for improving the generalization performance of modern neural networks. After success several traditional random data images (including flipping, translation, or rotation), recent surge interest in implicit techniques occurs to complement techniques. Implicit augments training samples feature space, rather than pixel resulting generation semantically meaningful data. Several on have been introduced classification tasks. However, few approaches regression tasks with continuous/structured labels, such as object pose estimation. Hence, we are motivated propose method semantic hand By considering distances poses, proposed implicitly generates extra space. We two additional improve this augmentation: metric learning and hand-dependent augmentation. Metric aims learn representations reflect distance For estimation, distribution augmented poses can be regulated by managing representations. Meanwhile, specifically designed estimation prevent meaningless from being generated (e.g., hands simple interpolation between both hands). Further, demonstrate effectiveness using datasets: STB RHD.