Condense loss: Exploiting vector magnitude during person Re-identification training process

Abstract

The magnitudes of features and weights significantly affect the gradients during the training process. L2 normalized softmax losses (such as NormFace, CosFace, ArcFace, etc.) and Naive softmax losses both reduce the magnitudes of image features in the training process and achieve good results in face recognition and person re-identification tasks, respectively. In this paper, we fully utilize the feature vector magnitudes and propose Condense loss for Re-ID tasks, which replaces the inner production of Naive softmax loss with the negative Euclidean distance. Condense loss generates negative radial gradients when updating weight parameters to push all features compacter. Because the coefficients of tangential gradients (the tangential component of the gradients) are related to feature magnitudes, it ideally provides monotonically decreasing tangential gradients, resulting in gradually diminishing updates that enhance the stability of the training process. We also introduce a margin parameter into Condense loss to enlarge inter-class distances and thus help the model learn more discriminative features. Mathematical analysis is given in this paper, and we have conducted sufficient experiments focusing on Re-ID tasks to prove the corresponding conclusion. The experimental results demonstrate that the Condense loss achieves competitive results compared to the state-of-the-art methods in the person re-identification task. At the same time, it also has a good performance in face recognition tasks.