ViLNM: Visual-Language Noise Modeling for Text-to-Image Person Retrieval

Text-to-image person retrieval (TPR) focuses on finding a specific person based on the textual description, and most methods implicitly assume the training image-text pairs are correctly aligned. In practice, the image-text pairs exist under-correlated or false-correlated due to the low quality of the images and annotation errors. Meanwhile, remarkable similarities between different person identities may lead to a mismatch between text and image. To tackle the two issues, we present a Visual-Language Noise Modeling (ViLNM) method that successfully captures robust cross-modal associations even with noise. Specifically, we design a Noise Token Aware (NTA) module that eliminates the words in the textual description that do not match the image, utilizing the matched words to establish a more reliable association. Besides, to enhance the recognition ability of the model for different person identities, we propose a Joint Inter and Intra-Modal Contrastive Loss (JII) and Local Aggregation (LA) module to increase the feature differences between different person identities. We conduct comprehensive experiments on three public benchmarks, and ViLNM performs best.