Abstract: RGB-based 3D hand pose estimation methods frequently produce physiologically invalid gestures due to depth ambiguity and self-occlusion. Existing methods typically adopt complex networks and a large amount of data to avoid invalid gestures by automatically mining the physical constraints of the hand. These networks exhibit high computational complexity and thus are difficult to be deployed into mobile devices. In consideration of this problem, a novel knowledge distillation framework, called Explicit Knowledge Distillation, is proposed to enhance the performance of small pose estimation networks. The proposed teacher network has interpretable knowledge, explicitly passing the physical constraints to the student network. Experimental results on three benchmark datasets with different sized models demonstrate the potential of our approach.