Text-to-Image Person Search Based on SSAN Model and Re-rank Post-Processing

Text-based person search (TBPS) is a task aiming to retrieve a target person image that most meet the given text description from a large image gallery. However, due to the huge modality gap, intra-class variance and limited data, it is very challenging to solve this fine-grained cross-modal retrieval task. Therefore, this paper proposes a method based on semantically self-aligned network (SSAN) and re-rank post-processing to solve above challenges. Firstly, in the training process, this paper adopts a more effective data augmentation strategy to solve the problem of limited data. Secondly, different from most TBPS tasks, this paper uses ResNet101 as the feature extractor of image branch to extract the global and fine-grained features of image more effectively. Thirdly, this paper adds a re-rank post-processing to improve the performance of person search. Extensive experiments demonstrate that the method proposed in this paper achieves a competitive performance in CUHK-PEDES dataset, and has been significantly improved over the previous models in Rank-1, Rank-5, Rank-10 and mAP.