Dual Compositional Learning in Interactive Image Retrieval

We present an approach named Dual Composition Network (DCNet) for interactive image retrieval that searches the best target a natural language query and reference image. To accomplish this task, existing methods have focused on learning composite representation of text to be as close embedding possible. refer Network. In work, we propose loop with Correction models difference between in space matches it query. That is, consider two cyclic directional mappings triplets (reference image, query, image) by using both also joint training loss can further improve robustness multimodal learning. evaluate proposed model three benchmark datasets retrieval: Fashion-IQ, Shoes, Fashion200K. Our experiments show our DCNet achieves new state-of-the-art performance all datasets, addition consistently improves multiple are solely based Moreover, ensemble won first place Fashion-IQ 2020 challenge held CVPR workshop.