Overview

LOOKALIKES is a fine-grained object-counting dataset that ships with our paper Just Say the Word: Annotation-Free Fine-Grained Object Counting — D’Alessandro, Mahdavi-Amiri, Hamarneh (2025).

What it’s for: stress-testing text-specified, class-agnostic counting models (i.e., zero-shot / prompt-shot / open-world, or whatever we’re calling it these days). The dataset is test-only and deliberately tricky: most images contain multiple visually similar sibling sub-categories (e.g., jalapeño vs. habanero, penne vs. gnocchi), and the task is to count only the instances of the requested sub-category given a text prompt.

If you’ve got a pre-trained counter, this is a drop-in benchmark for measuring generalization under fine-grained confusion. Run your model with a sub-category prompt, report per-class counts, and let us know how it holds up.

(P.S. don’t forget to cite our work!)

1) Download

The dataset can be directly downloaded from the following link:

2) What’s inside

Data-LookAlikes/
├── <Broad-Category>/                         
│   ├── images/
│   │   ├── <fname>                 # JPG/PNG
│   │   └── ...
│   └── ...
├── annotations.all.testset         # all labels
├── usage_example.ipynb             # a demo for loading images
├── LICENSE.md
└── README.md

Dataset stats

• Images: 1,037
• Domains: peppers, apples, pasta, potatoes, tomatoes, waterfowl, wildflowers, gemstone finishes
• Sub-categories: 27
• Annotation type: point labels per instance (+ per-subcategory counts)
• Avg density: 36.4 objs/img (max 656)
• Multi-class: 83% of images contain ≥2 sibling sub-categories

3) Annotation schema

annotations.all.testset.json:

{
  "<broad-category>": {
    "sub_categories": ["<sub-cat>", ...],
    "data": [
        {"image": {"fname": "<fname>", "width": "W", "height": "H"},
         "annotations": {"<sub-cat>": [{"point": [x, y]}, ...], ...}
        },
        ...
    ]
  }
}

4) Evaluation protocol

Task: predict the count for each sub-category present in an image. Primary metrics: In our paper, we calculate MAE / RMSE per sub-category and macro-averaged across all.

There are helper functions in usage_example.ipynb that can be used for macro-averaging as follows:

annotations = json.load(open("./annotations.all.testset.json"))
mae_across_splits = []
for broad_cat, sub_cat in  get_subcategory_keys(annotations):
    subcat_split = get_subcategory_split(annotations, broad_cat, sub_cat)
    per_example_subcat_mae = []
    for example in sub_cat_split:
        # calculate MAE
        mae = None # add your logic
        per_example_subcat_mae.append(mae)
    # calculate mae for sub_cat
    mae_across_splits(np.mean(per_example_subcat_mae)

print(np.mean(mae_across_splits))

5) License (summary)

• Images: original creators’ licenses (Flickr/Unsplash/Pexels/…); see metadata.

• Annotations: custom research & individual use license.

  • ✅ Academic & personal use
  • ❌ Corporate, commercial, governmental, policing, military, surveillance uses

Full text: see LICENSE.md.

6) Citation

@article{dalessandro2025justsaytheword,
  title   = {Just Say the Word: Annotation-Free Fine-Grained Object Counting},
  author  = {D'Alessandro, Adrian and Mahdavi-Amiri, Ali and Hamarneh, Ghassan},
  journal = {arXiv preprint arXiv:2504.11705},
  year    = {2025}
}