abidlabs/ORCA-sample

ORCA 100-image random sample

A random sample of 100 images (with their annotations) drawn from the ORCA dataset (WongYukKwan/ORCA), the benchmark from ORCA: Object Recognition and Comprehension for Archiving Marine Species (WACV 2026, arXiv:2512.21150).

How it was sampled

• 100 images sampled uniformly at random with a fixed seed (random.Random(42)) from the 14,645 images in the source dataset.
• The 100 sampled images span all 670 species categories in expectation; the sample contains 269 bounding-box annotations (with captions).
• Annotations in data.json are filtered to only the sampled images; the full categories taxonomy (670 entries) is kept unchanged so category IDs still resolve.

Structure

COCO format, identical to the source dataset:

• data.json — {"images": [...], "annotations": [...], "categories": [...]}
• images/ — the 100 sampled image files

Each annotation carries bbox, caption, and label (0 = LLM-generated positive caption, 1 = LLM-generated negative caption, 2 = expert-refined positive caption).

License

The source dataset is released under CC-BY-4.0; this sample inherits that license. Please cite the original ORCA paper if you use this sample:

@InProceedings{Wong_2026_WACV,
    author    = {Wong, Yuk-Kwan and Liang, Haixin and Ma, Zeyu and Chen, Yiwei and Zheng, Ziqiang and Gotama, Rinaldi and Sebastian, Pascal and Sparks, Lauren D. and Yeung, Sai-Kit},
    title     = {ORCA: Object Recognition and Comprehension for Archiving Marine Species},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month     = {March},
    year      = {2026},
    pages     = {1597-1609}
}