sivakornchong/fd_widerface_yolov8

Training for educational purposes

Face Detection and Age Classification Demonstration

Sivakorn (Oak) Chongfeungprinya

Purpose

In this scenario, we assume there is an application from health promotion board that would propose different health exercises to different age groups. Hence, a face detection and an age classification algorithm would be very helpful for determining the user's age.

Age range definition based on Health Promotion Board of Singapore

• (A0) Young Children: 0-6 years
• (A1) Children and Youth: 7-17 years
• (A2) Youth Adults: 18-25 years
• (A3) Adults: 26-49 years
• (A4) Older adults: 50 years

Disclaimer: This is a model training demonstration for educational purposes only.

Algorithm and Training Data

Face Detection Model:

• YOLOV8 base architecture
• Finetuned based on publicly available WiderFace dataset
• Eventual MAP(50) on test dataset is: 64% just after 20 epochs.
• Below is a sample of testing results

Age Classification Model:

• RESNET34 base architecture
• Finetuned based on publicly available UTKFace dataset. The dataset is crtopped using the engine from face detection model.
• Model is trained up to epoch 4, which is selected to prevent overfitting on training data
• Weighted average accuracy is 74% on the test dataset.
• Below is an example of training dataset.

Deployment

• A simple .py script is developed to deploy both models in a sample demo.
  • The face detection model is used to detect a face. That face is cropped as an image for further processing.
  • The classification model is used for each cropped image.
  • The results are then demonstrated on screen.
• Demo

Potential Improvement

• Improve variability within the dataset used for training both models.
  • For classification model specifically, data preprocessing could be done to rotate the faces around and improve the accuracy upon deployment.
  • Train with more epochs for detection model.
• Model quantization to reduce model sizes for deployment on smaller devices (phones).
• Data for application (webcam images) could slightly be different from the faces from internet used for classification model training. It would be good to expand the training dataset into the actual application sets.

References

• YOLOV8: https://github.com/ultralytics/ultralytics
• RESNET34: https://pytorch.org/vision/main/models/generated/torchvision.models.resnet34.html#torchvision.models.resnet34
• UTKface dataset: https://susanqq.github.io/UTKFace/
• WiderFace dataset: http://shuoyang1213.me/WIDERFACE/