FaceViT: A lightweight multitask Vision Transformer for face detection, age prediction and gender classification

FaceViT: A small-sized multi-task Vision Transformer trained from scratch for face detection, age estimation, and gender prediction, demonstrating the ability of the Vision Transformer to perform great across different tasks such as object detection and multiple classifications simultaneously.

Dataset

For this project I have used the UTK Faces dataset, which you can download with a Kaggle account from here:

• UTK Faces dataset

NOTE: This dataset is heavily imbalanced when it comes to ages of people, while also containing a lot of relatively low quality images. This could be a limiting factor for performance over the different tasks.

Some predictions

Training time loss and metrics

Loss plot

Metrics plot

NOTE: During this training experiment no augmentations were used.

Evaluation metrics on validation set

| Metric Name | Value | Training Epochs | |-----------------------|---------|--------| | Top-3 Age Accuracy | 61 % | 43 | | Face Bounding Box MSE | 0.0054 | 43 | | Gender Accuracy | 75 % | 43 |