billpsomas/icir

i-CIR Dataset (Hugging Face)

website | arxiv | github

About

i-CIR (Instance-Level Composed Image Retrieval) is a curated benchmark for composed image retrieval where each instance corresponds to a specific, visually indistinguishable object (e.g., a particular landmark). Each query combines an image of the instance with a text modification, and retrieval is evaluated against a database containing rich hard negatives (visual / textual / compositional).

Key stats

• Instances: 202
• Total images: ~750K
• Composed queries: 1,883
• Avg database size / query: ~3.7K images
• Includes challenging hard negatives per instance.

---

Dataset Structure

On Hugging Face, i-CIR is hosted as WebDataset shards for scalable/robust downloads and streaming.

icir/
├── webdataset/
│   ├── query/
│   │   ├── query-000000.tar
│   │   ├── query-000001.tar
│   │   └── ...
│   └── database/
│       ├── database-000000.tar
│       ├── database-000001.tar
│       └── ...
├── annotations/
│   ├── query_files.csv
│   ├── database_files.csv
├── VERSION.txt
└── LICENSE

---

Annotations format

• query_files.csv: each row is (image_path, text_query, instance_id)
• database_files.csv: each row is (image_path, text_query, instance_id) (the text field may be unused for database features depending on the pipeline)

Inside each WebDataset sample, we store:

• an image (.jpg/.png/...)
• a json payload with: img_path, text, instance

---

Download

One-liner download (recommended):

pip install -U huggingface_hub
huggingface-cli download billpsomas/icir --repo-type dataset --local-dir ./data/icir --revision main

Python (equivalent):

from huggingface_hub import snapshot_download
snapshot_download(repo_id="billpsomas/icir", repo_type="dataset", local_dir="./data/icir", revision="main")

---

Using the dataset (feature extraction)

You can extract features directly from the WebDataset shards (no image folder extraction needed):

python3 create_features.py \
  --dataset icir \
  --icir_source wds \
  --icir_wds_root ./data/icir \
  --backbone clip \
  --batch 512 \
  --gpu 0

---

License

The dataset is released under CC BY-NC-SA 4.0. Please see LICENSE for details.

---

Citation

If you use i-CIR in your research, please cite:

@inproceedings{
    psomas2025instancelevel,
    title={Instance-Level Composed Image Retrieval},
    author={Bill Psomas and George Retsinas and Nikos Efthymiadis and Panagiotis Filntisis and Yannis Avrithis and Petros Maragos and Ondrej Chum and Giorgos Tolias},
    booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
    year={2025}
}