Confidence-adapted meta-interaction for unsupervised person re-identification

Most unsupervised person re-identification (ReID) approaches combine clustering-based pseudo-label prediction with feature learning, and perform the two steps in an alternating fashion for training ReID models. However, incorrect/noisy pseudo-labels are often present due to various variations (e.g., human pose, illumination, and viewpoint, etc.). Such noisy pseudo-labels may harm the trained ReID models. In order to use diverse variations/information while minimizing negative influence of the noisy pseudo-labels, we propose a confidence-adapted meta-interaction (CAMI) method by explicitly exploring the interaction between the believable supervision (reliable pseudo-labels) and the diverse information. Specifically, CAMI iteratively trains the ReID model in a meta-learning manner, in which the training images are dynamically divided into a reliable set and an unreliable set. At each iteration, the pseudo-labels of images are predicted by clustering and the training images are divided by the proposed confidence-adapted sample disentanglement (CASD) method. To adapt the changes of the pseudo-labels and gradually refine the division, the CASD method dynamically predicts the pseudo-label confidence. It divides the training images into the reliable set (with high confidence pseudo-labels) and the unreliable set (with low confidence pseudo-labels), respectively. Then a meta-interaction method is proposed for training the ReID model, which consists of a meta-training step to use the believable supervision of the reliable set and a meta-testing step to use the diverse information of the unreliable set. Meanwhile, a bridge model is dynamically built to refine the unreliable set based on the believable supervision from the reliable set. The CAMI is evaluated by two unsupervised person ReID settings, including the image-based and the video-based. The experimental results on four datasets demonstrate the superiority of the proposed CAMI.