xuzhaopan/Real-CIRR_Real-FashionIQ

Real-CIRR / Real-FashionIQ

Overview

This repository provides Real-CIRR and Real-FashionIQ, two real-world noisy datasets for Composed Image Retrieval (CIR). They are constructed based on the original CIRR and FashionIQ benchmarks, and introduce realistic web noise through Google multisearch.

Compared with synthetic noise generated by random shuffling, these datasets are designed to better reflect practical CIR scenarios, where mismatches may arise from the reference image, modification text, or retrieved target image.

---

Dataset Characteristics

• Real-CIRR

  • Built upon CIRR
  • Represents fine-grained general-domain CIR
  • Approximately 75K training triplets

• Real-FashionIQ

  • Built upon FashionIQ
  • Represents fashion-domain CIR
  • Approximately 85K training triplets

These two datasets preserve the task structure of standard CIR benchmarks while introducing more realistic noisy correspondence from web retrieval, making them suitable for studying robust CIR under real-world noise.

---

Files

This repository currently provides the following resources:

• Real-CIRR_train.json Training annotations for Real-CIRR

• Real-CIRR_target_image.zip Target images used as supervision in Real-CIRR

• Real-FashionIQ_train.json Training annotations for Real-FashionIQ

• Real-FashionIQ_target_image.zip Target images used as supervision in Real-FashionIQ

---

How to Use

Step 1: Download the original datasets

Please first download the original CIRR and FashionIQ datasets from their official sources.

These original datasets provide:

• the reference images
• the modification texts
• the standard CIR data structure

Step 2: Download this dataset

Download the files in this repository.

Step 3: Construct the training data

Use:

• the reference image and modification text from the original CIRR/FashionIQ
• the target image provided in Real-CIRR / Real-FashionIQ

That is, the training triplet is formed as:

[ \langle I_r, T_m, I_t^{real} \rangle ]

where:

• (I_r): reference image from the original CIR benchmark
• (T_m): modification text from the original CIR benchmark
• (I_t^{real}): target image provided by this dataset

In other words, this dataset is intended to be used by keeping the original query side unchanged and replacing the target image with the released real-web target image as the training supervision signal.

---

Intended Use

This dataset is designed for:

• robust composed image retrieval
• noisy correspondence learning
• evaluation under real-world web noise
• comparison with synthetic-noise training settings

---

Notes

• The dataset is released for academic research only.
• It is built upon existing CIR benchmarks and should be used together with the original CIRR and FashionIQ datasets.
• The target images in this repository serve as the supervision signal for training under realistic noisy correspondence.

---

Citation

If you use this dataset in your research, please cite our paper.

@article{xu2026realcirr,
  title={Noisy Correspondence Decomposition for Robust Composed Image Retrieval},
  author={Zhaopan Xua, Pengfei Zhou, Chenrui Zhou, Xi Chen, Wangbo Zhao, Panpan Zhang, Jianning Zhang, Xiaojiang Peng, Hongxun Yaoa, and Kaipeng Zhang},
  journal={},
  year={2026}
}