CIRCLED: A Multi-turn CIR Dataset with Consistent Dialogues across Domains

The CIRCLED dataset has been introduced as a significant advancement in Multi-Turn Composed Image Retrieval (MTCIR), addressing previous limitations in dialogue-history consistency and domain restrictions by expanding upon existing datasets like FashionIQ, CIRR, and CIRCO. This dataset comprises 22,608 multi-turn sessions, enhancing both scale and generality compared to its predecessors.

This development is crucial as it provides a high-quality benchmark for future research in multi-turn image retrieval, facilitating improved accuracy and consistency in user interactions across various domains.

The introduction of CIRCLED aligns with ongoing efforts in the field to enhance image retrieval systems, particularly in addressing challenges such as ambiguity in user queries and the need for more sophisticated retrieval frameworks, as seen in recent advancements like DeliCIR and MCoT-MVS, which also aim to refine the accuracy and efficiency of image retrieval processes.