Joint Weights-Averaged and Feature-Separated Learning for Person Re-identification

Although existing research has made considerable progress in person re-identification (re-id), it remains challenges due to intra-class variations across different cameras and the lack of cross-view paired training data. Recently, there are increasing studies focusing on using generative model to augment training samples. One of the main obstacles is how to use generated unlabeled samples. To address this issue, we propose a joint learning framework without label predictions, including re-id learning and data generation end-to-end. Each person encodes into a feature code and a structure code. The generative module is able to generate cross-id images by decoding structure code with switched feature code. For generated unlabeled data, we average re-id model weights instead of label predictions. Moreover, we expand the distance between inter-class feature code. Our approach improves the accuracy of the re-id model and the quality of the generated data. Experiments on several benchmark datasets shows that our method achieves competitive results with most state-of-the-art methods.