salvacarrion/face-detection

Curated Face Detection Dataset

A drop-in dataset for training classical face detectors (Viola-Jones-style cascades, Haar-cascade classifiers, or any sliding-window pipeline that needs small grayscale face crops + a large pool of natural-image negatives).

The face crops come from four well-known source datasets, all preprocessed to grayscale (single channel) with consistent square cropping. The hard-negative source is Caltech-256 with face/people/human categories filtered out, kept as raw color JPGs so the user can extract negatives at any resolution.

Dataset structure

Three splits are available:

Split	Rows	Contents
train	~68 k	CelebA 48×48 faces + FDDB 48×48 faces + CBCL 19×19 train faces/nofaces
test	~24 k	CBCL 19×19 test faces + nofaces (classic VJ benchmark)
negatives	~30 k	Caltech-256 filtered color JPGs (bootstrap negative pool)

Each row has four columns:

Column	Type	Description
image	Image	decoded image (grayscale for CBCL/CelebA/FDDB, color for Caltech)
label	ClassLabel	0 = noface, 1 = face
source	string	"celeba" | "fddb" | "cbcl" | "caltech"
category	string | null	Caltech category folder name (e.g. "001.ak47"); null for all other sources

Local layout (development)

datasets/
├── README.md                        # This file (HF dataset card)
├── STATS.md                         # Per-source counts, mean/std, mosaics
├── celeba/train/                    # 50,000 PNGs @ 48×48 grayscale + meta.json
├── fddb/train/                      # 11,383 PNGs @ 48×48 grayscale + meta.json
├── cbcl/
│   ├── train/
│   │   ├── faces/                   #  2,429 PNGs @ 19×19 grayscale + meta.json
│   │   └── nofaces/                 #  4,548 PNGs @ 19×19 grayscale + meta.json
│   └── test/
│       ├── faces/                   #    472 PNGs @ 19×19 grayscale + meta.json
│       └── nofaces/                 # 23,573 PNGs @ 19×19 grayscale + meta.json
└── caltech/
    ├── README.md
    └── source/                      # 254 categories, ~29,900 color JPGs (~1.1 GB)
        ├── meta.json
        └── NNN.<category>/*.jpg

Recommended use per source

Pick one face source for your training set; treat CBCL train as a historical reference, not a useful training set. CBCL test stays valuable as an academic benchmark for direct comparison with the literature.

Source	Faces / Negatives	Recommended for	Notes
CelebA	50,000 / —	Training (default) — large, easy to scale	Frontal headshots, very tight alignment, narrow pose distribution
FDDB	11,383 / —	Training when pose variation matters	Merged train+valid; varied poses, lighting, occlusion
CBCL train	2,429 / 4,548	Not recommended for training	Too small and narrow, pre-cropped 19×19
CBCL test	472 / 23,573	Test set for academic comparison	Standard VJ benchmark; hostile 1:50 imbalance
Caltech-256	— / ~30k images	Negative pool (hard-neg mining)	Raw color JPGs, faces/people/humans filtered

Crop conventions for face PNGs

FDDB and CelebA use a tight square crop with 10% margin; CBCL is copied verbatim from the MIT release.

• FDDB: bbox xywh from COCO annotations, filtered to min(w, h) >= 24 px (smaller boxes upsample poorly to 48×48). Margin 10%, square crop centered on the bbox center, reflect-padded if it falls outside the source image.
• CelebA aligned: face bbox computed from the 5-point landmarks. With the eye and mouth midpoints, face_height ≈ eye_to_mouth_distance × 2.63 (anthropometric ratio: eye-to-mouth is ~0.38 of full face height). Center is shifted up from the eye-mouth midpoint by 0.37 × eye_to_mouth.
• CBCL: pre-cropped by MIT, no margin, no upsampling, native 19×19.

Each meta.json records the source raw image, original bbox/landmarks, and extraction params, so any PNG is reproducible from the original raw datasets.

Caltech-256 raw

Three categories are filtered out because they contain human/face content that would leak positives into a "negatives" set:

• 253.faces-easy-101 (frontal faces)
• 159.people (whole-body people)
• 112.human-skeleton (skulls/skeletal structure)

Filter rule: any directory whose name contains face, people, or human (case-insensitive). The full kept/excluded category list lives in caltech/source/meta.json.

The raw color JPGs are kept as-is — the user picks resolution and pool size when extracting hard-negative patches with their training pipeline.

How to use

With the datasets library (recommended)

from datasets import load_dataset

# All splits at once
ds = load_dataset("salvacarrion/face-detection")

# Individual splits
train     = load_dataset("salvacarrion/face-detection", split="train")
test      = load_dataset("salvacarrion/face-detection", split="test")
negatives = load_dataset("salvacarrion/face-detection", split="negatives")

# Filter by source
celeba = train.filter(lambda x: x["source"] == "celeba")
fddb   = train.filter(lambda x: x["source"] == "fddb")

# Access an image (returns a PIL Image)
img = train[0]["image"]

With this training repo

After downloading the dataset, build NumPy bundles ready for cascade training:

# 1) Download
huggingface-cli download salvacarrion/face-detection \
    --repo-type dataset --local-dir datasets

# 2) Build training-ready NPY bundles (one face source + Caltech negative pool)
python tools/prepare_data.py \
    --face-source celeba \
    --n-faces 10000 \
    --resolution 24 \
    --augment

# 3) Train your cascade
python main.py train --dataset-path datasets ...

Sources and licenses

This dataset is a derivative work combining four sources, each under their own terms. Use is restricted to non-commercial research.

Source	Original URL	License
FDDB (Roboflow re-pack v1, 2022)	https://public.roboflow.ai/object-detection/undefined	CC BY 4.0
CelebA (aligned)	http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html	Research-only, non-commercial
MIT CBCL Face Database #1	http://cbcl.mit.edu/software-datasets/FaceData2.html	Research-only, non-commercial
Caltech-256	https://data.caltech.edu/records/nyy15-4j048	Research-only, non-commercial

By using this dataset you agree to abide by the licenses of all four underlying sources. CelebA and CBCL explicitly prohibit redistribution; this re-pack is provided for research use only under that understanding. If you represent any of the original dataset providers and want this taken down, open an issue on the dataset repo.

Cite the original sources, not this re-pack, in academic work.