Knowledge transfer between networks and its application on gait recognition

Convolutional neural network (CNN) has achieved promising results in many vision tasks. However, a deep CNN model requires long training process and consumes large amounts of storage space. In this paper we propose a novel framework to boost the speed and reduce the size of CNN based models, and to test it on the human gait recognition task. The idea is to develop a small and fast fully connected network (FCN) which retains the learning ability of the original large CNN. Specially, we build a large CNN model, and train with gait gallery data to obtain parameters of each layer. Then we design a small FCN that inherit the softmax weight matrix from the CNN, and use the gallery data to generate parameters of rest of the layers for the FCN. We use CASIA Gait Dataset B to evaluate the proposed framework, and test the performance under multiple covariate factors. The experimental results suggest that the extended FCN is able to retain most of the learning abilities from the large CNN model. Comparing with CNN, the extended FCN has reduced size with significant speed boost.