Person re-identification (Re-ID) systems match people across different cameras, but their visual representations often degrade when deployed in new environments. The authors of ReText argue that natural language descriptions are more stable across domains than raw visual features, because captions are less affected by lighting and other domain-specific biases. To exploit this, they propose a training method that injects textual supervision into an image-based Re-ID model.
Because standard multi-camera Re-ID datasets do not come with text annotations, the authors first use a fine-tuned vision-language model to generate captions for those datasets. ReText then trains the model on three tasks at once: the original Re-ID objective, image-text matching, and text-guided image reconstruction. The method also incorporates single-camera data, and unlike prior work that used such data only as images, ReText uses the textual annotations that naturally accompany them.
Experiments on cross-domain benchmarks show that ReText significantly outperforms state-of-the-art methods, supporting the claim that text boosts generalization. The paper, accepted for WACV 2027, makes the expanded multi-camera datasets and the VLM weights publicly available.