This project focuses on a vision model that classifies Social Media Users based on their age and gender (male/female) using their profile picture as input. Initially, we implemented a simple CNN-based model that provided accurate results by pairing age and gender, which improved accuracy significantly. However, to surpass the state-of-the-art models in gender accuracy, which stands at approximately 70-75% on the UTK Face dataset, we adopted a transformer-based model.
Specifically, we utilized the Mobile Vision Transformer (Mobile ViT), which employs a CNN architecture as an encoder, making the model slightly smaller and faster than conventional transformers.
We have achieved an accuracy of 76.67%, surpassing the state-of-the-art model for classifying age. This outcome demonstrates the effectiveness of combining age and gender for the classification task.