Composed Image Retrieval

Composed image retrieval (CIR) aims to retrieve images matching a multimodal query consisting of a reference image and a text description specifying desired modifications. Current research heavily focuses on zero-shot CIR, developing methods that avoid the need for expensive triplet-labeled datasets, often employing techniques like textual inversion, contrastive learning, and large language models to generate synthetic training data or improve feature representation. This field is significant for its potential to enhance image search capabilities beyond keyword-based systems, impacting applications in digital humanities, remote sensing, and e-commerce by enabling more nuanced and flexible image querying.