Frame Correlation Knowledge Distillation for Gait Recognition in the Wild

Guozhen Peng, Shaoxiong Zhang, Yuwei Zhao, Annan Li, Yunhong Wang

Abstract: Recently, large deep models have achieved significant progress on gait recognition in the wild. However, such models come with a high cost of runtime and computational resource consumption. In this paper, we investigate knowledge distillation (KD) for gait recognition, which trains compact student networks by using a cumbersome teacher network. We propose a novel scheme, named Frame Correlation KD (FCKD), to transfer the frame correlation map (FCM) from the teacher network to the student network. Since the teacher network usually learns more frame correlations, transferring such FCM from teacher to student makes the student more informative and mimic the teacher better, thus improving the recognition accuracy. Extensive experiments demonstrate the effectiveness of our approach in improving the performance of compact networks.