Relaxed Pairwise Learned Metric for Person Re-Identification Martin Hirzer, Peter M. Roth, Martin K¨ ostinger, and Horst Bischof Institute for Computer Graphics and Vision Graz University of Technology Abstract. Matching persons across non-overlapping cameras is a rather challenging task. Thus, successful methods often build on complex fea- ture representations or sophisticated learners. A recent trend to tackle this problem is to use metric learning to find a suitable space for match- ing samples from different cameras. However, most of these approaches ignore the transition from one camera to the other. In this paper, we propose to learn a metric from pairs of samples from different cameras. In this way, even less sophisticated features describing color and texture information are sufficient for finally getting state-of-the-art classification results. Moreover, once the metric has been learned, only linear pro- jections are necessary at search time, where a simple nearest neighbor classification is performed. The approach is demonstrated on three pub- licly available datasets of different complexity, where it can be seen that state-of-the-art results can be obtained at much lower computational costs. 1 Introduction Person re-identification, i.e., recognizing an individual across spatially disjoint cameras, is becoming one of the major challenges in visual surveillance. Typical applications include but are not limited to tracking criminals, analyzing crowd movements in public places, and finding children who lost their parents. Since the number of public areas that become subject to video surveillance is ever growing, efficient, automatic systems are required to reduce the load on human operators. In general, person re-identification is very challenging for several reasons. First, the appearance of an individual can vary extremely across a network of cameras due to changing view points, illumination, different poses, etc. Second, there is a potentially high number of “similar” persons (e.g., people wear rather dark clothes in winter). Third, in contrast to similar large scale search problems typically no accurate temporal and spatial constraints can be exploited to ease the task. Thus, motivated by the high number of practical applications and still unresolved difficulties there has been an increased scientific interest (e.g., [1– 10]) in recent years, and also various benchmark datasets [11, 8, 6] have been published.