A unified perspective of classification-based loss and distance-based loss for cross-view gait recognition. (May 2022)

• Record Type: Journal Article
• Authors: Han, Feng; Li, Xuejian; Zhao, Jian; Shen, Furao
• Is Part Of: Pattern recognition. Volume 125(2022)
• Publication Date: 2022-05
• Subjects: Biometrics -- Gait recognition -- Computer vision -- Metric learning -- Angular softmax loss function -- Triplet loss function; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2021.108519
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We propose a new loss function that guides the network to map the gait feature to a separable yet discriminative feature space. Our loss function utilizes A-Softmax to learn a separable feature in the cosine space and triplet loss to increase the distance between feature vectors of different subjects and decrease the distance between the feature vectors of the same subjects. A-Softmax and triplet loss are optimized in different spaces. In order to make the training process feasible, we add a batch-normalization layer after extracting gait feature (before the last fully-connected layer) to reduce the impact of optimizing two different losses. We conduct comprehensive experiments on CASIA-B dataset and TUM GAID dataset. The experiment results show that using our loss function with GaitSet as our backbone network exceeds the previous state-of-the-art performance under the same experiment settings. Abstract: Gait can be used to recognize people in an uncooperative and noninvasive manner and it is hard to imitate or counterfeit, which makes it suitable for video surveillance. The current solutions for gait recognition are still not robust to handle the conditions when the view angles of the gallery and query are different. We improve the performance of cross-view gait recognition from the perspective of metric learning. Specifically, we propose to use angular softmax loss to impose an angular margin for extracting separable features. At the same time, we use triplet loss to make the extracted features more discriminative. Additionally, we add a batch-normalization layer after extracting gait features to effectively optimize two different losses. We evaluate our approach on two widely-used gait dataset: CASIA-B dataset and TUM GAID dataset. The experiment results show that our approach outperforms the prior state-of-the-art approaches, which shows the effectiveness of our approach.