Abstract:
Gait recognition is an important biometric technology that identifies a person by using walking posture. Recently, most gait recognition methods either take the human gait as a whole to generate Global Feature Representations (GFR) or equivalently divide the human gait into multiple local regions to establish Local Feature Representations (LFR). However, we observe that LFR or GFR does not adequately represent the human gait because that LFR only focuses on the detailed information of each local region and GFR pays more attention to the global context information. On the other hand, the partition manner of the local regions is fixed, which only focuses on the local information of several specific regions. Motivated by this observation, we propose a novel mask-based network, named GaitMask, for gait recognition. GaitMask is built based on the Mask-based Local Augmentation (MLA), which is used to learn more comprehensive feature representations. MLA is a dual-branch structure consisting of a GFR extraction as the trunk and a mask-based LFR extraction as the branch. Specifically, the mask-based LFR extraction consists of a pair of complementary masks, where one mask randomly drops a region of the input feature maps and the other one only preserves this region. The complementary mask can be used to generate more comprehensive LFR and enhances the robustness of feature representations of the trunk. Experiments on two popular datasets demonstrate that our method achieves state-of-the-art results. Specifically, the proposed method significantly increases the performance in complex environments.
