Label smoothing decoupling and gradient suppression for person re-identification

An effective person re-identification (ReID) model needs to be discriminative to distinguish similar people in the same scene, and generalizable to identify the same person in different scenes. But for the commonly used cross-entropy loss and triplet loss, the gradient will always be back-propagated no matter how easy the sample pair is, which causes the network to focus too much on unnecessary identification of easy sample pairs. To solve this problem, we improve loss functions with gradient suppression of easy sample pairs. We measure whether a sample pair is easy or not by the angle of the vectors, as the similarity is completely determined by the angle. For any given sample pair, we first decompose the gradient of the feature or the class center along its angle and amplitude directions. If the sample pair is an easy one, then the angle component of the gradient will be suppressed, while the amplitude component will be still preserved. These two components can then synthesize a new gradient. This decomposition-suppression-synthesis process can be applied to both the cross-entropy loss and the triplet loss. Also, we analyze the different effects of label smoothing on class centers and target features. Since the proposed method is independent of the network structure, it can be applied to almost all ReID networks. Compared with the ReID baseline using Resnet50 as the backbone, the proposed method can increase mAP from 76.4% to 77.5% on DukeMTMC, and from 85.9% to 87.1% on Market1501.