theoriclabs/face-or-not

Face or Not

theoriclabs/face-or-not is a balanced binary image-classification dataset of 4,000 128x128 RGB crops. The label answers one narrow question: does this crop contain an Open Images Human face annotation?

This is classification, not face localization, identification, recognition, or biometric matching. It has no names or identity labels.

Splits

split	no_face	face	total
train	1,600	1,600	3,200
validation	200	200	400
test	200	200	400

Each Open Images source image contributes at most one row, and source image IDs do not cross splits. The test labels are intended for one final held-out evaluation; model selection should use the validation split.

Source and licenses

All image pixels and labels derive from the official Open Images V7 release. The Open Images Human face class ID is /m/0dzct.

• Open Images lists the source images as Creative Commons Attribution 2.0 (CC BY 2.0). The builder accepts only metadata rows whose license URL exactly matches that license.
• Google licenses the Open Images annotations under CC BY 4.0.
• Every example retains its Open Images image ID, original image and landing URLs, author and author profile, title, license URL, annotation source, and crop coordinates. These fields preserve attribution after cropping.

Open Images states that it tried to identify CC BY images but gives no warranty about each image's license status. The per-row source and attribution fields are therefore part of the dataset and must be preserved. Users who need additional assurance should verify the linked original before redistribution or production use.

Suggested attribution: “Derived from Open Images V7 annotations, © Google LLC, CC BY 4.0; source images and their named authors are listed per row under CC BY 2.0.”

How the rows were selected

The deterministic builder uses the 41,620-image Open Images validation split, whose annotations are human verified.

For face, it selects the largest /m/0dzct box in each source image, rejects group boxes, depictions, and boxes smaller than one percent of the source image, then adds 25% context around a square crop. Occluded and edge-truncated faces are kept. For no_face, it uses only images with an explicit human-verified negative /m/0dzct label and takes the largest centered square. Both crops are resized to 128x128 with Lanczos resampling.

Candidates are ranked by SHA-256 of a versioned seed, label, and source image ID. Fixed per-class counts are assigned to train, validation, and test. There is no random state or dependence on upstream CSV row order.

Fields

• image: 128x128 RGB image
• label: no_face (0) or face (1)
• source_image_id: Open Images ID, unique across the dataset
• source_image_url: official CVDF/Open Images copy used by the builder
• original_url, original_landing_url: original Flickr links
• license: per-image license URL
• author, author_profile_url, title: attribution metadata
• crop_norm_json: actual normalized [xmin, ymin, xmax, ymax] crop
• crop_strategy: face-box-plus-25pct or center-square
• annotation_source: Open Images box method or human-verified-negative

Reproduce

prepare_dataset.py pins SHA-256 hashes for the four official metadata files. It generates the full manifest before downloading any pixels.

Version-one manifest SHA-256: 8026d6deb5fcf6f9d9945c996d73a7dc69322ff466b05269591cd95d3105e8fa.

python prepare_dataset.py

# Small, balanced build: 4 examples per class in each split.
uv run --with 'datasets>=3,<5' --with 'Pillow>=10,<13' \
  python prepare_dataset.py --build --smoke-per-class 4 \
  --output-dir face-or-not-smoke

# Full local build. This does not publish anything.
uv run --with 'datasets>=3,<5' --with 'Pillow>=10,<13' \
  python prepare_dataset.py --build --output-dir face-or-not-dataset

Publishing is deliberately separate and requires both flags:

uv run --with 'datasets>=3,<5' --with 'Pillow>=10,<13' \
  python prepare_dataset.py --build --output-dir face-or-not-dataset \
  --push-repo theoriclabs/face-or-not --confirm-push

Intended use and limitations

The dataset is sized for a small tutorial classifier, not a production safety or biometric system. It inherits Open Images sampling and annotation biases. The positive class is face-centered by construction, negatives are center crops, and some faces are occluded or truncated. The IsDepiction filter is imperfect: a manual smoke review found face-shaped dolls and masks among rows annotated as Human face. A high test score therefore does not establish performance on surveillance footage, tiny faces in wide scenes, drawings, masks, unusual camera spectra, or demographic subgroups.

The images may contain people and other personal information even though the dataset contains no identity labels. Do not use it to identify people or infer sensitive traits. Respect removal requests and the linked source licenses.