---
license: cc-by-4.0
task_categories:
  - object-detection
  - keypoint-detection
language:
  - en
pretty_name: SoloFace2
size_categories:
  - 100K<n<1M
tags:
  - face-detection
  - facial-landmarks
  - tinyML
  - embedded-AI
  - single-face
  - resource-constrained
  - bounding-box
  - keypoints
  - WIDER-FACE
  - COCO
annotations_creators:
  - machine-generated
multilinguality:
  - monolingual
---

# SoloFace2: A Large-Scale Single-Face Dataset with Landmarks

SoloFace2 is a large-scale face detection and landmark regression dataset
optimized for **resource-constrained deployment** (TinyML, edge AI, microcontrollers).
Every image contains **exactly one visible human face or none at all**,
making it ideal for training lightweight single-face detectors.

## Key Features

- **5 facial landmarks** per face (eyes, nose, mouth corners)
- **Single-face only** — no multi-face ambiguity
- **Stratified splits** — 80/10/10 train/val/test, stratified by face presence
- **5× augmentation** (WIDER FACE, SoloFace train) — geometric + color + crop
- **224×224 JPEGs** — ready-to-train, quality 95
- **Multiple sources** for diversity: studio portraits, natural scenes, crowds, backgrounds

## Dataset Statistics

| Split | Face (p=1) | No-face (p=0) | Total |
|-------|-----------|---------------|-------|
| Train | 46,689 | 67,147 | 113,836 |
| Val | 5,792 | 8,397 | 14,189 |
| Test | 5,909 | 8,387 | 14,296 |
| **Total** | **58,390** | **83,931** | **142,321** |

*Face:No-face = 1:1.44 overall (41% face). Training uses 50/50 batch sampling.*

## Source Composition

| Source | Generator | Face | No-face | Description |
|--------|-----------|------|---------|-------------|
| SoloFace train | RetinaFace | 27,498 | 28,862 | COCO-derived, 5× augmented |
| SoloFace test | YuNet | 1,669 | 180 | COCO-derived, raw images |
| SoloFace val | YuNet | 196 | 21 | COCO-derived, raw images |
| WIDER FACE train | RetinaFace | 22,255 | 0 | Natural scenes, 5× augmented |
| WIDER FACE val | RetinaFace | 5,175 | 0 | Natural scenes, 5× augmented |
| COCO no-person | YuNet | 1,597 | 54,868 | Person-free scenes |
| **Total** | | **58,390** | **83,931** | |

## Label Format

Each image has a corresponding row in `labels.csv`:

| Column | Type | Description |
|--------|------|-------------|
| `image` | str | JPEG filename |
| `p` | int | 1 = face present, 0 = no face |
| `x` | int | Bbox left edge (224×224 pixels) |
| `y` | int | Bbox top edge |
| `w` | int | Bbox width |
| `h` | int | Bbox height |
| `right_eye_x`, `right_eye_y` | int | Right eye center |
| `left_eye_x`, `left_eye_y` | int | Left eye center |
| `nose_x`, `nose_y` | int | Nose tip |
| `right_mouth_x`, `right_mouth_y` | int | Right mouth corner |
| `left_mouth_x`, `left_mouth_y` | int | Left mouth corner |
| `score` | float | Detection confidence (0.0 when p=0) |

**Coordinate space:** All integers are in 224×224 pixel coordinates.
Divide by 224 to normalize to [0, 1].

**Landmark order (subject's perspective):**
0. Right eye | 1. Left eye | 2. Nose tip | 3. Right mouth | 4. Left mouth

## Face Detection Models

Labels were generated by two complementary face detectors:

| Model | Usage | Output |
|-------|-------|--------|
| **RetinaFace** (insightface) | WIDER FACE, SoloFace train | Bbox + 5 landmarks |
| **YuNet** (OpenCV) | SoloFace test/val, COCO | Bbox + 5 landmarks |

Both models target the same 5 anatomical points. The column order is
normalized to YuNet convention in the CSV.

## Data Augmentation (5×)

Applied to WIDER FACE and SoloFace training images (matching original SoloFace):

| Augmentation | Parameters |
|-------------|------------|
| Rotation | ±15° random |
| Scaling | ±20% random |
| Horizontal flip | 50% probability |
| Brightness | ±30% random |
| Contrast | ±30% random |
| Random crop | Up to 10% from edges |

All augmentations preserve bounding box and landmark consistency.
Augmented variants are named `*_aug0.jpg` through `*_aug4.jpg`.

## Splits

Train/val/test splits are 80/10/10, stratified by the `p` column.
The `splits.csv` file maps each image to its split.

```
split_dataset() using sklearn.model_selection.train_test_split
  stratify on column p
  random_state = 42
  train : val : test = 80 : 10 : 10
```

## Download & Extract

Images are split across 5 tar.gz archives (~640 MB each, 3.2 GB total). VGGFace2 images are excluded:

```
images_part_01.tar.gz   images_part_03.tar.gz   images_part_05.tar.gz
images_part_02.tar.gz   images_part_04.tar.gz
```

**Option 1: Extract script (recommended)**
```bash
# Download all images_part_*.tar.gz files, then:
python extract_images.py
```

**Option 2: Manual extraction**
```bash
# Linux/Mac
for f in images_part_*.tar.gz; do tar -xzf "$f"; done

# Windows (PowerShell)
Get-ChildItem images_part_*.tar.gz | ForEach-Object { tar -xzf $_ }
```

## File Structure

```
soloface2/
├── README.md              ← This dataset card
├── images/                ← 142,321 JPEG images (224×224, quality 95)
├── images_part_01.tar.gz  ← Image archive chunk 1
├── ...
├── images_part_05.tar.gz  ← Image archive chunk 5
├── extract_images.py      ← Archive extraction script
├── labels.csv             ← All labels (image, p, bbox, landmarks, score)
├── splits.csv             ← Train/val/test split assignments
└── generate_splits.py     ← Split generation script (reproducibility)
```

## Intended Use

1. Training lightweight face detection models for edge deployment
2. Facial landmark regression on resource-constrained hardware
3. Single-face detection benchmarking
4. TinyML model compression and quantization research
5. Privacy-preserving on-device face detection

## Usage

```python
import pandas as pd
from pathlib import Path
from PIL import Image

# Load labels and splits
labels = pd.read_csv("labels.csv")
splits = pd.read_csv("splits.csv")
df = labels.merge(splits, on="image")

# Training subset
train = df[df["split"] == "train"]
row = train.iloc[0]
img = Image.open(f"images/{row['image']}")  # 224×224 RGB

# Normalize coordinates to [0, 1]
DATA_SIZE = 224
bbox = [row["x"] / DATA_SIZE, row["y"] / DATA_SIZE,
        row["w"] / DATA_SIZE, row["h"] / DATA_SIZE]
kps = [
    (row["right_eye_x"] / DATA_SIZE, row["right_eye_y"] / DATA_SIZE),
    (row["left_eye_x"] / DATA_SIZE, row["left_eye_y"] / DATA_SIZE),
    (row["nose_x"] / DATA_SIZE, row["nose_y"] / DATA_SIZE),
    (row["right_mouth_x"] / DATA_SIZE, row["right_mouth_y"] / DATA_SIZE),
    (row["left_mouth_x"] / DATA_SIZE, row["left_mouth_y"] / DATA_SIZE),
]
```

## License

| Source | License |
|--------|---------|
| VGGFace2 | Research use only |
| SoloFace | CC BY 4.0 |
| WIDER FACE | Research use |
| COCO 2017 | CC BY 4.0 |

This dataset is for **research purposes only**. Ensure compliance with
VGGFace2 and WIDER FACE terms before commercial use.

## Citation

```bibtex
@dataset{soloface2,
  title     = {SoloFace2: A Large-Scale Single-Face Dataset with Landmarks},
  author    = {Saha, Bidyut and Samanta, Riya},
  year      = {2026},
  publisher = {Hugging Face},
  note      = {Extended from SoloFace (10.5281/zenodo.14474899) with
               WIDER FACE and RetinaFace landmarks}
}
```

## Contact

- Bidyut Saha: sahabidyut999@gmail.com
- Riya Samanta: study.riya1792@gmail.com