Incremental XQDA Metric Learning for Person Reidentification

Person re-identification is a useful technique to automatically match observations of the same person cross different times and different camera views. It attracts extensive attention and researches in computer vision community because of its application and challenge. An effect way to tackle the problem is to learn a useful distance metric from training examples. Then the learned metric could be used for distance calculations between a probe image and images from the gallery. For a real application, the labeled training samples usually increase gradually along the time. To keep the performance of a system for re-identification tasks, the learned model needs to be updated according to the newly added training sets. Although the learned model can be retrained with the whole dataset, the procedure is usually time-consuming. In this paper, we propose an incremental metric learning method based on the widely used XQDA metric for person re-identification. The key idea is that the covariance matrices of similar and dissimilar example pairs can be effectively updated incrementally in the XQDA metric learning algorithm. Then the final metric could be deduced in an incremental way. The proposed approach method is validated on two public available datasets, and the experimental results indicate the effectiveness of our proposed approach.