Contrastive completing learning for practical text–image person ReID: Robuster and cheaper

Abstract

Text–image person re-identification (TIReID) seeks to leverage textual descriptions for the retrieval of target pedestrians. Due to its versatility, TIReID has gained increasing attention. However, manual annotation of textual descriptions and identity labels can be time-consuming and costly, limiting its scalability in practical settings. Privacy concerns and poor data storage can lead to data loss or ineffectiveness, further exacerbating challenges in real-world scenarios. To address these limitations, we propose for the first time incomplete Text–image person re-identification (iTIReID), which comprises a small amount of complete pairwise data and a large amount of incomplete data, where all identity labels are unavailable. We introduce a novel Contrastive Completing Learning (CCL) framework for iTIReID, consisting of two stages: Pure Contrastive Learning (PCL) and Feature Completion Contrastive Learning (FCCL). In PCL, only complete pairwise data is utilized for training, which serves as a preliminary improvement of the model’s capacity and prepares for the upcoming feature completion stage. In FCCL, available features are used to complete missing modality features and facilitate effective training with incomplete data. During this process, Cross-modal Semantic Measure (CSM) is proposed to leverage intra-modality similarity to measure cross-modal similarity and filter out features with the highest semantic similarity, thereby circumventing modality discrepancy. Semantic-Weighted Generation (SWG) is proposed to generate approximate features based on the semantic similarity weight of the similar features. To fully leverage pairwise data for label-free training, we introduce the contrastive CMPM (CCMPM) loss for contrastive learning to achieve weakly supervised training. Experimental results verify the effectiveness of our proposed methods and demonstrate competitive performance compared to fully supervised methods using complete data.