Face to Body Mass Index

This model predicts Body Mass Index (BMI) with one image of a human face, with state-of-the-art results.

Performance

The performance of this model is better than the benchmarks set by state-of-the-art method on VisualBMI dataset as of Jan 2024 by 39.5%.

Original dataset

After training 10 epoches, the model has a MAE loss of 3.45 on test dataset.

Augmented dataset

After training 7 epoches, the model has a MAE loss of 3.02 on test dataset.

Installation

1. Clone this repository by running:

git clone [email protected]:liujie-zheng/face-to-bmi-vit.git
cd face-to-bmi-vit

2. Install conda here.
3. Depending on your operating system, install dependencies by running:

conda env create -f environment_linux.yml
conda activate face2bmi

or

conda env create -f environment_mac.yml
conda activate face2bmi

Run a demo in terminal

1. (Optional) replace ./data/test_pic.jpg with your own image. Note: for your own image, a face should occupy a substantial part of the image for optimal results.
2. In root directory, run:

cd scripts
conda run -n face2bmi --no-capture-output python demo.py

if you encounter a PermissionError: [Errno 13] Permission denied error, instead run:

sudo conda run -n face2bmi --no-capture-output python demo.py

Train it by yourself

In root directory, train the original unaugmented dataset by running:

cd scripts
conda run -n face2bmi --no-capture-output python run.py

or train the augmented dataset by running:

cd scripts
conda run -n face2bmi --no-capture-output python run.py --augmented=True