Face Detection with Age & Gender Prediction

A deep learning pipeline that detects faces in images, video, and real-time webcam streams, then predicts the age group and gender of each detected person.

Deep Learning 2022 — Final Project.

Overview

The system is a two-stage pipeline:

1. Face detection — a YOLOv5 model locates faces and returns bounding boxes.
2. Age & gender classification — each detected face is cropped and passed to a CNN classifier that predicts an age group and the gender.

The final notebook ties both stages together and runs on still images, video files, and live webcam input.

Dataset

• Face detection: trained on the WIDER FACE dataset (converted to YOLO format) and evaluated against COCO 2017 face annotations.
• Age & gender classification: trained on the UTKFace dataset, whose filenames encode the age and gender of each subject.

Age is discretised into 9 groups of ten years each (00-10, 11-20, …, 81-90).

Model & Approach

Face detection

• YOLOv5 (yolov5s), fine-tuned on WIDER FACE at 1280 px input resolution.
• Exported to ONNX for inference.

Age & gender classification

Two backbones were trained and compared, each with a final layer producing 9 age classes + 2 gender classes:

• ResNet-18 (ImageNet-pretrained) — used as the faster classifier in the final pipeline.
• InceptionResNetV1 (facenet-pytorch, VGGFace2-pretrained).

Training details:

• Loss: MSE on the age head + BCE on the gender head (summed).
• Optimizer: Adam (lr=0.001, amsgrad, weight_decay=1e-6) with ReduceLROnPlateau scheduling.
• 25 epochs, batch size 64, 80/10/10 train/validation/test split.

Results

Evaluated on the UTKFace test split:

Model	Age (mean group error)	Gender accuracy
ResNet-18	~0.60	~93%
InceptionResNetV1	~0.60	~91%

Age error is the mean absolute difference between the predicted and true age-group index (each group spans 10 years).

Example detections:

Repository Contents

Notebook	Description
Face Detection Model Creation.ipynb	Prepares the WIDER FACE data and trains/configures the YOLOv5 detector.
Face Extraction.ipynb	Visualises the detector output, returning cropped face images.
Age & Gender Prediction Model Creation.ipynb	Trains and evaluates the ResNet-18 and InceptionResNetV1 classifiers.
Face Detection: Age & Gender Detection.ipynb	Final pipeline combining detection and classification for images, video, and webcam.

Other directories:

• Models/ — trained model weights (YOLOv5, ResNet-18, InceptionResNetV1).
• DATA/ — sample test images (WIDER_TEST) and rendered results (Results).
• Figures/ — training/validation loss plots.

Running the Project

The notebooks were developed in Google Colab with GPU acceleration and support a local fallback (they use the current working directory when Colab is unavailable).

1. Clone the repository.
2. Open the notebook of interest in Jupyter or Colab.
3. To run the full pipeline, open Face Detection: Age & Gender Detection.ipynb; it loads the pretrained weights from Models/ and runs inference on the sample images in DATA/WIDER_TEST/.
4. To retrain, use the two Model Creation notebooks (a GPU is strongly recommended).

Requirements

• Python 3
• PyTorch & TorchVision
• YOLOv5 (loaded via torch.hub)
• facenet-pytorch
• OpenCV, Pillow, NumPy, Matplotlib
• albumentations, pybboxes (data preparation)

Authors

• Joan Gracia
• Pere Fraga
• Josep Ricci

License

Released under the GNU General Public License v3.0. See LICENSE.