Learning Joint Gait Representation Via Quintuplet Loss Minimization

Gait recognition is an important biometric technique relevant to video surveillance, where the task is to identify people at a distance by their walking patterns captured in the video. Most of the current approaches for gait recognition either use a pair of gait images to form a cross-gait representation or rely on a single gait image for unique-gait representation. These two types of representations empirically complement one another. In this paper, we propose a new Joint Unique-gait and Cross-gait Network (JUCNet) representation, to combine the advantages of both schemes, leading to significantly improved performance. A second contribution of this work is a tailored quintuplet loss function, which simultaneously boosts inter-class differences by pushing different subjects further apart and contracts intra-class variations by pulling same subjects closer. Extensive tests demonstrate that our method achieves the best performance tested on multiple standard benchmarks, compared with other state-of-the-art methods.