Parallel CNN Classification for Human Gait Identification with Optimal Cross Data-set Transfer Learning

Abstract

Convolutional neural networks (CNN) have been used in multi-scene and cross-view gait recognition. In practice, new dataset and application scenarios continue to emerge, and how to bridge the new to the prior experience is an important optimization problem. Transfer learning has been successfully used to map the model from the source domain to the target domain in many image classification tasks. However, in human gait identification, models are designed to be more complex to improve the feature extraction abilities and spatial-temporal information utilization. GaitSet, for instance, has a global parallel pipeline to collect various-level set information. In these cases, how to fine-tune the model in the target domain does not have a determinate answer. In this paper, we investigate the impact of each layer on parallel-structured CNN. To be specific, we gradually freeze the parameters of GaitSet from its higher layers to the lower and see the fine-tuning performance. We find that such a parallel structure is more robust to the co-adaption phenomenon compared with the single pipeline. Moreover, although the increase of frozen layers has an adverse effect on the accuracy, it can reach the maximum efficiency with only one convolution layer unfrozen.