Boosting Text-To-Image Person Re-Identification With Generative Hard Negative

Text-to-image person re-identification(TIReID) aims to find matching target person from the textual descriptions. Due to the inter-modal gap, existing methods cannot effectively distinguish the similar images of different identities corresponding to the same description. To solve this problem, we propose GHNR, GAN-based Hard Negative Re-identification Framework, to utilize data expansion to improve the performance of TIReID. Specifically, we design a joint training framework that combines a cross-modal generative adversarial network(GAN) and a TIReID network, using the GAN to obtain cross-modal images and provide them as negative samples to the TIReID network to improve the ability to discriminate the subtle differences of similar samples. To ensure the GAN has the correct generative direction, we design the Cross-modal Loss (CML) to update the generator and provide new generative data at each training epoch. To better utilize the expanded data, we design the Triplet Joint Alignment Loss (TJA). Experiments on three datasets demonstrate the superiority and effectiveness of our method.