Evaluating Features for Person Re-Identification

Many features have been proposed for person re-identification, but most of them are the combination of several different kinds of single features. And there is no research about what role the single features play and which can be fused together to improve the performance. In this paper, we explore eight single features (four colors, two textures, two gradients) and their fusions. Evaluations are conducted on four public datasets with two metric learning algorithms. Experimental results show that the color features are the more effective features comparing with texture and gradient features. It can greatly improve the accuracy when the single features are fused with different type, region segmentation and quantization. The conclusion can guide us to fuse several single features to improve the performance.