Predicting Attractiveness using Computer Vision

Traditional Machine Learning Approach

Feature Generation

The features computation part of the pipeline requires the location of facial landmars of the input images. These landmarks can be generated by the CLM-framework. I have already included the landmarks localized using this framework in the data directory of this repo, and you can directly work with them.

Example of extracting facial features:

Dimensionality Reduction, ML Models, and Evaluation

python trainModel -model linear_model -featuredim 20

The -featuredim argument specifies the number of components chosen by PCA which are how many dimensions to be reduced.

After PCA, the -model argument is used to indicate the traditional machine learning models including Support Vector Machines (svm), Random Forests (rf), and Gaussian Process Regression (gpr). Checkout the source to change hyperparameters and other options.

Applying Pearson Correlation for result analysis with learner's prediction. For example, figure below shows PC for evaluation of Linear Regression:

Deep Learning Approach:

Details of the implementation can be found in the paper SCUT-FBP: A Benchmark Dataset for Facial Beauty Perception, the result of the creation and research on the dataset.

Model:

Network architecture of our CNN for facial beauty prediction:

Results + Example Usage:

Sample visualized results from the prediction of models for audiences to compare. The blue block represents the human's rating and the green block represents the model's rating

These pictures are celebrities in Vietnamese showbiz community and how their facial beauty were rated by our model.

What's included

face-rating/
├── data/
│   ├── ratings.txt
│   ├── landmarks.txt
│   ├── features_ALL.txt
├── source/
|   ├── machine_learning/    
│       ├── generateFeatures.py
│       ├── trainModel.py
│       ├── cross_validation.py
|   ├── deep_learning/    
│       ├── build_model.py
│       ├── utils.py
│       ├── face_rating.ipynb
|   ├── deeplearning_result.ipynb
|   ├── traditional_result.ipynb

Requirements

• Python 2.7
• numpy
• pandas
• scikit-learn
• keras
• tensorflow

Dataset

The SCUT-FBP dataset has been used. Please cite their research if you happen to use this dataset. The facial landmarks computer on this particular dataset are available in the data/ directory.

License

MIT