CoCo-IR: Contextual Composed Image Retrieval
CoCo-IR extends composed image retrieval from single-turn instructions to multi-turn interactions, allowing users to progressively refine visual search intent. Its LMM-based context-aware reasoner reads the full interaction history and produces Transformable Image Embeddings (TIE) that evolve across turns. The authors also introduce an autonomous data engine that generates contextual retrieval examples and mines hard negatives through model-guided verification. The reported results are 39.4 mAP@5 on CIRCO and 44.1 R@1 on four-turn CoCo-IR dialogues, compared with 28.2 R@1 for existing methods on the latter setting.
Why it's worth reading
Multi-turn visual search is becoming a concrete retrieval setting rather than a single-query benchmark. CoCo-IR is timely because it contributes both a dialogue benchmark and a context-aware model, with reported four-turn results that expose weaknesses in prior systems.