sharmithas151005/face-detection-widerface-yolo-style

Face Detection Model (YOLO-Style, Trained on WIDER FACE)

This repository provides a YOLO-style face detection model trained on the WIDER FACE dataset.
The model predicts bounding boxes across a fixed spatial grid, using S = 20 and B = 3 bounding boxes per grid cell.

---

Model Details

Feature	Value
Input Resolution	320 × 320
Grid Size (S)	20 × 20
Boxes per Cell (B)	3
Output Format	(cx, cy, w, h, confidence)
Framework	PyTorch

• (cx, cy) are normalized center coordinates
• (w, h) are normalized width and height
• confidence is the face likelihood score

---

Installation

git clone https://huggingface.co/sharmithas151005/face-detection-widerface-yolo-style
cd face-detection-widerface-yolo-style
pip install -r requirements.txt

Dataset: WIDER FACE

Download and prepare:

WIDER/ ├─ WIDER_train/images/ ├─ WIDER_val/images/ └─ wider_face_split/wider_face_train_bbx_gt.txt

Update dataset paths inside dataset.py.

Training

python train.py
--images_dir "path/to/WIDER_train/images"
--annotation_file "path/to/wider_face_train_bbx_gt.txt"
--epochs 50
--batch_size 16
--lr 0.0001

This will save trained weights as model.pth.

Inference (Run Face Detection)

python inference.py --image path/to/image.jpg --weights model.pth

This will display the image with detected faces.