Circle-Based Ratio Loss for Person Reidentification

Figure 1

The diagrammatic explanation of the circle-based ratio loss. For simplicity, the feature dimension is set as 2 so the hypersphere can be represented as a circle on the 2D plane. The features and corresponding classification weights of three classes are denoted with the dots and solid lines in three different colors, respectively. d and D represent the maximal intraclass distance and the minimal interclass distance, respectively. By minimizing the ratio of d and D, not only the intraclass distance of the blue class is contracted but also the interclass distance of the blue class and green class is enlarged. The ratio loss can improve the distribution of feature space effectively and help the model learn discriminative features.