Rethinking Composed Image Retrieval Evaluation: A Fine-Grained Benchmark from Image Editing
This addresses the evaluation gap in CIR for researchers and practitioners by providing a more rigorous benchmark, though it is incremental as it builds on existing benchmarks.
The authors tackled the problem of evaluating composed image retrieval (CIR) by creating a fine-grained benchmark called EDIR from image editing, which includes 5,000 queries across categories, and found that state-of-the-art models like RzenEmbed and GME struggle to perform consistently across all subcategories.
Composed Image Retrieval (CIR) is a pivotal and complex task in multimodal understanding. Current CIR benchmarks typically feature limited query categories and fail to capture the diverse requirements of real-world scenarios. To bridge this evaluation gap, we leverage image editing to achieve precise control over modification types and content, enabling a pipeline for synthesizing queries across a broad spectrum of categories. Using this pipeline, we construct EDIR, a novel fine-grained CIR benchmark. EDIR encompasses 5,000 high-quality queries structured across five main categories and fifteen subcategories. Our comprehensive evaluation of 13 multimodal embedding models reveals a significant capability gap; even state-of-the-art models (e.g., RzenEmbed and GME) struggle to perform consistently across all subcategories, highlighting the rigorous nature of our benchmark. Through comparative analysis, we further uncover inherent limitations in existing benchmarks, such as modality biases and insufficient categorical coverage. Furthermore, an in-domain training experiment demonstrates the feasibility of our benchmark. This experiment clarifies the task challenges by distinguishing between categories that are solvable with targeted data and those that expose intrinsic limitations of current model architectures.