README.md

---
license: other
license_name: cc-by-nc-4.0-mixed-rights
license_link: LICENSE
task_categories:
- image-classification
language:
- en
tags:
- image
- datasets
- webdataset
- computer-vision
- image-retrieval
- animal-re-identification
- cat
- fine-grained-recognition
pretty_name: Individual Cats in the Wild
size_categories:
- 10K<n<100K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.tar
  - split: validation
    path: data/validation-*.tar
  - split: test
    path: data/test-*.tar
---

# Individual Cats in the Wild (ICW)

ICW is an identity-disjoint benchmark for individual cat identification in
unconstrained images. It contains **82,791 JPEG images of 19,877 individual
cats**, collected from six public adoption and rescue platforms and curated for
fine-grained recognition and image retrieval research.

ICW accompanies the MeowID project:

> **MeowID: A Dual-Expert Retrieval System for Individual Cat Identification**
>
> Zhangchi Hu, Yi Shang, Haocheng Yang, Qiwei Hu, and Yuzheng Li (2026)

- Project repository: https://github.com/RicePasteM/MeowID
- Model weights: https://huggingface.co/RicePasteM/MeowID-Base

## Dataset structure

The release uses identity-preserving [WebDataset](https://github.com/webdataset/webdataset)
TAR shards. Shards target approximately 1 GiB, and all images of one identity
remain in the same shard.

```text
ICW/
├── data/
│   ├── train-00000.tar
│   ├── ...
│   ├── train-00023.tar
│   ├── validation-00000.tar
│   └── test-00000.tar
├── cats.csv
├── metadata.csv
├── splits.csv
├── manifest.json
├── verify_webdataset.py
└── extract_to_imagefolder.py
```

| Split | Identities | Images | Shards | Images per identity |
| --- | ---: | ---: | ---: | ---: |
| Train | 18,877 | 77,094 | 24 | 3–19 (mean 4.08) |
| Validation | 500 | 2,851 | 1 | 5–14 (mean 5.70) |
| Test | 500 | 2,846 | 1 | 5–17 (mean 5.69) |
| **Total** | **19,877** | **82,791** | **26** | **3–19 (mean 4.17)** |

Identity sets are strictly disjoint across the three splits.

## Sample format

Each WebDataset example contains adjacent members with the same key:

```text
00001234_000001.jpg
00001234_000001.json
```

The JSON member contains the identity, split, original path, normalized crop
box, source provenance, and the corresponding row from `metadata.csv`.
`identity_id` is the benchmark label; `image_id` identifies an observation of
that individual.

## Loading

Install the vision dependencies:

```bash
pip install "datasets[vision]"
```

The dataset card defines all three splits, so it can be streamed directly:

```python
from datasets import load_dataset

dataset = load_dataset("RicePasteM/ICW", streaming=True)
sample = next(iter(dataset["train"]))

image = sample["jpg"]
metadata = sample["json"]
identity_id = metadata["identity_id"]
```

An explicit WebDataset configuration is also possible:

```python
from datasets import load_dataset

files = {
    "train": "hf://datasets/RicePasteM/ICW/data/train-*.tar",
    "validation": "hf://datasets/RicePasteM/ICW/data/validation-*.tar",
    "test": "hf://datasets/RicePasteM/ICW/data/test-*.tar",
}
dataset = load_dataset("webdataset", data_files=files, streaming=True)
```

To restore the conventional image-folder layout:

```bash
python extract_to_imagefolder.py /path/to/ICW /path/to/icw_imagefolder
```

This produces `train/<identity>/<image>.jpg`,
`validation/<identity>/<image>.jpg`, and `test/<identity>/<image>.jpg`.

## Metadata

All tables are UTF-8 encoded. The original `cats.csv` and `metadata.csv` use a
UTF-8 byte-order mark, so pass `encoding="utf-8-sig"` when reading them.

### `cats.csv`

One row per identity. It includes the eight-digit `cat_folder` label, source
platform, source-side animal identifier, source-reported attributes, profile
URL, and preserved source metadata.

### `metadata.csv`

One row per image. It includes `cat_folder`, `image_filename`, source
provenance, source-reported attributes, original URLs, normalized crop box,
curation assignment, and preserved source metadata.

### `splits.csv`

One row per identity with its split, image count, and containing shard. This
table is the fastest way to map an identity to a TAR file without scanning the
archives.

### `manifest.json`

The release manifest records sample and identity counts, byte sizes, SHA-256
checksums, and key ranges for every shard. Run the included verifier after a
download:

```bash
python verify_webdataset.py /path/to/ICW
```

Source-reported fields may be incomplete, outdated, or inaccurate and should
not be treated as verified biological labels.

## Construction

The release was built through the following stages:

1. Public adoption and rescue profiles were collected from six platforms.
2. Automatic filtering retained cat images and profiles with sufficient views.
3. Cats were segmented, the exterior background was masked, and crops were
   tightened around the subject.
4. Near-duplicate observations and low-quality samples were removed through
   automated comparison and manual review.
5. Identities were assigned to mutually exclusive train, validation, and test
   splits.

| Source | Identities |
| --- | ---: |
| PetFinder 2026 | 16,785 |
| PetRescue | 2,881 |
| 汪汪喵呜孤儿院 | 115 |
| RSPCA Animals | 47 |
| BC SPCA | 41 |
| Verhuisdieren | 8 |

Source names describe provenance, not endorsement of this dataset or its
authors.

## Intended uses

ICW is intended for non-commercial research and evaluation involving
individual animal identification, fine-grained image retrieval, route-aware
face and whole-animal recognition, and representation learning under changes
in pose, viewpoint, lighting, background, camera, and occlusion.

The dataset is not intended for identifying people, inferring pet ownership,
contacting source organizations or individuals, making automated
animal-welfare decisions, or commercial deployment.

## Limitations and responsible use

- The dataset is drawn from adoption and rescue platforms rather than a random
  sample of the global cat population.
- Geographic, platform, breed, age, health, photographic, and curation biases
  may affect model behavior.
- Listings and URLs can become outdated after collection.
- Free-text source metadata can contain contact details or other incidental
  information. Do not use it to identify, profile, or contact people.
- Image availability on a public webpage does not by itself imply that the
  image is freely relicensable.

Users are responsible for evaluating legal, ethical, privacy, and source-site
requirements in their jurisdiction and application.

## Data quality

The release was validated on 17 August 2026:

- 19,877 identities and 82,791 paired JPEG/JSON examples;
- no identity overlap between train, validation, and test;
- no identity split across multiple shards;
- deterministic member ordering and valid sidecar JSON;
- byte sizes and SHA-256 checksums verified for all 26 shards;
- all original metadata rows matched to exactly one image.

## Licensing and source rights

The dataset-specific selection, identity organization, split assignments,
curation annotations, and original documentation contributed by the MeowID
authors are licensed under the **Creative Commons Attribution-NonCommercial
4.0 International license (CC BY-NC 4.0)**. See [`LICENSE`](LICENSE).

This license applies only to material for which the MeowID authors hold the
necessary rights. **Third-party photographs, source listing text, trademarks,
and linked content are excluded from that license** and remain subject to their
respective copyright holders' rights and source-platform terms. No endorsement
by any source platform is implied.

To request removal of an identity, image, source record, or stale contact
information, open an issue in the
[MeowID repository](https://github.com/RicePasteM/MeowID/issues) and include
only the minimum path or source identifier needed to locate the record.

## Citation

```bibtex
@article{hu2026meowid,
  title  = {MeowID: A Dual-Expert Retrieval System for Individual Cat Identification},
  author = {Hu, Zhangchi and Shang, Yi and Yang, Haocheng and Hu, Qiwei and Li, Yuzheng},
  year   = {2026}
}
```