---
license: other
license_name: research-only-mixed
pretty_name: Face Detection (CelebA + CBCL + Caltech-256)
task_categories:
  - object-detection
  - image-classification
tags:
  - face-detection
  - viola-jones
size_categories:
  - 100K<n<1M
---

# Face Detection Dataset

Small grayscale face crops + a large pool of natural-image negatives, packaged
for classical face detectors (Viola-Jones, Haar cascades, sliding-window
classifiers).

## At a glance

| Split | Rows | Faces / Non-faces |
|---|---|---|
| `train`     | 106,977 | 102,429 / 4,548 |
| `test`      | 24,045 | 472 / 23,573 (CBCL benchmark) |
| `negatives` | 29,879 | — / 29,879 (Caltech-256) |

Same schema across all splits: `image`, `label` (0/1), `source`, `category`.

## Quick start

```python
from datasets import load_dataset

ds = load_dataset("salvacarrion/face-detection")

# All faces for training
faces = ds["train"].filter(lambda x: x["label"] == 1)

# CBCL benchmark for evaluation
test = ds["test"]

# Hard-negative pool (raw color JPGs at native resolution)
negs = ds["negatives"]
```

## Face sources (train split)

CelebA ships in **two paired variants at the same index** — same person,
two crop styles:

- `celeba` — loose portrait framing (hair and jaw visible).
- `celeba_aligned` — tight Viola-Jones crop (eyes pinned to fixed pixel
  positions, face fills the 48×48 frame).

CBCL is already canonically aligned, so it has a single variant.

| `source` | Label | Count | Notes |
|---|---|---|---|
| `celeba`         | 1 | 50,000 | Loose portrait. Filtered for frontal pose using the manual CelebA landmarks. |
| `celeba_aligned` | 1 | 50,000 | Same 50,000 faces tightly warped to CBCL-matched geometry. |
| `cbcl`           | 1 | 2,429 | MIT CBCL #1, upsampled from native 19×19. Already canonically aligned. |
| `cbcl`           | 0 | 4,548 | CBCL non-faces — matched-domain seed for stage 1 of cascade training. |

Picking `source == "cbcl"` from train returns **both faces and non-faces** —
filter by `label` to pick.

## Test split

The classic CBCL Viola-Jones benchmark, untouched. Faces and non-faces live
together in one split so accuracy is one pass.

## Negatives split

Caltech-256 source JPGs (native color, varying resolution). Categories
containing `face`, `people`, or `human` are excluded. Sample patches at
whatever resolution your detector needs.

## Aligned geometry (the `*_aligned` variants)

Two-point similarity transform — face shape preserved, no stretching:

- Left eye  → (12, 10) in a 48×48 frame
- Right eye → (36, 10)

The canonical positions were measured empirically from the mean of CBCL
training faces, so models trained on `celeba_aligned` generalize cleanly to
the CBCL test benchmark.

## License

Research / non-commercial use only. Cite the original sources:

- [CelebA](http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html)
- [MIT CBCL Face Database #1](http://cbcl.mit.edu/software-datasets/FaceData2.html)
- [Caltech-256](https://data.caltech.edu/records/nyy15-4j048)