CIRCUS

CIRCUS (Composed Image Retrieval Cleaned of Unimodal Shortcuts) is a metadata-only diagnostic resource accompanying Do Composed Image Retrieval Benchmarks Require Multimodal Composition? It audits CIRR, FashionIQ, LaSCo, and CIRCO for unimodal shortcuts and validates a subset of shortcut-free queries.

Loading

from datasets import load_dataset

validated = load_dataset(
    "EdinburghNLP/CIRCUS",
    "cirr",
    split="validated",
    revision="v1.0",
)
shortcut_free = load_dataset(
    "EdinburghNLP/CIRCUS",
    "cirr",
    split="shortcut_free",
    revision="v1.0",
)

Available configurations are cirr, fashioniq, lasco, and circo. Each configuration has four splits:

• full: all original benchmark evaluation queries with automatic audit labels.
• shortcut_free: all queries not solved by any evaluated unimodal configuration at K=10; this includes both composition-required and unresolved queries (CIRCUS-SF).
• human_audited: shortcut-free queries inspected during manual validity annotation, including both valid and rejected instances.
• validated: the human-audited queries judged to be valid CIR instances and retained in CIRCUSV.

Dataset	Full	Shortcut-free	Human-audited	Validated
CIRR	4,170	685	685	303
FashionIQ	6,003	4,069	2,000	586
LaSCo	30,031	18,418	2,000	758
CIRCO	220	56	56	42
Total	40,424	23,228	4,741	1,689

For FashionIQ and LaSCo, human validation uses stratified samples of 1,000 composition-required and 1,000 unresolved queries per dataset. Their validated splits are therefore validated diagnostic subsets, not complete cleanings of the full shortcut-free residues.

Images and source datasets

This repository does not contain or redistribute images. Image paths are portable identifiers relative to the corresponding source dataset. Users must obtain CIRR, FashionIQ, LaSCo, and CIRCO from their official sources and comply with each dataset's terms.

The source datasets use different licenses and terms, including MIT-style terms for CIRR and LaSCo annotations, the Community Data License Agreement for FashionIQ, and CC BY-NC 4.0 for CIRCO. Underlying NLVR2 and COCO images remain subject to their original terms. The other license marker reflects this mixed provenance; it is not a relicensing of source data.

Audit labels

At audit_k=10, a query is labelled shortcut when at least one evaluated unimodal configuration retrieves an annotated positive in the top 10. A shortcut-free query is composition_required when a multimodal configuration succeeds, and unresolved when no evaluated multimodal configuration succeeds. These are operational benchmark labels and do not imply that shortcut-labelled queries are invalid.

Limitations

CIRCUSV is intended as a diagnostic of multimodal dependence, not a complete replacement for the original benchmarks. Human validation is complete for the CIRR and CIRCO shortcut-free residues but sampled for FashionIQ and LaSCo. The resource is metadata-only and requires separately downloaded source images.

Citation

@inproceedings{attimonelli2026circus,
  title={Do Composed Image Retrieval Benchmarks Require Multimodal Composition?},
  author={Attimonelli, Matteo and De Bellis, Alessandro and Gema, Aryo Pradipta and Saxena, Rohit and Sekoyan, Monica and Kwan, Wai-Chung and Pomo, Claudio and Suglia, Alessandro and Jannach, Dietmar and Di Noia, Tommaso and Minervini, Pasquale},
  booktitle={Advances in Neural Information Processing Systems},
  year={2026}
}

Code and additional documentation: https://github.com/EdinburghNLP/CIRCUS