Source-free domain adaptive person search

Person search aims to locate and identify individuals within given scene images. Despite the emergence of numerous studies in this area, few have addressed the issue of domain shift. Some researchers have acknowledged this gap and studied person search in an unsupervised domain adaptation setting. Assuming that the labeled source data is accessible during the adaptation process, they improve generalization on the target domain by aligning source and target representations. However, access to source domain data is often restricted due to privacy protection or data proprietary concerns. In this paper, we investigate person search under a source-free domain adaptation setting, seeking to generalize the source-trained person search model to the target domain without accessing the source data. To this end, we propose a novel source-free domain adaptive person search (SFPS) method. On the one hand, we introduce hard pseudo label learning and soft relational consistency learning to effectively distill target domain information into the source-trained model, employing the mean-teacher framework. On the other hand, we propose instance-level discriminative feature learning to enhance pedestrian feature representations in the target domain from both cross-model and intra-model perspectives. Extensive experiments demonstrate the superiority of our SFPS in adapting a source-trained model to the target domain. Furthermore, experimental results using sketches instead of photos as probes indicate that our SFPS can be readily applied to images of other modalities, achieving commendable generalization performance.