beauty scale

face ratings
face ratings
when no face is found

Inspiration

Well over the past few months I researched a lot to find any ML Model which is able to predict their score but unfortunately didn't get one. So then I decided to take a step ahead and create own model which will predict the beauty.

What it does

When the lens detects any face it triggers the ML component and it captures the face texture of the face which is the nearest one as input and then further the input is served to ML model to predict its beauty label with percentage of correctness.

How we built it

Initially we had a template for multi-class classification. The dataset had to be of proper format as required.The training and building took almost 12 hours with accuracy of 80%.Further the files generated from the python script and copied to the resources of the project section. We are using Post-training Quantized model and is used as input to the ML component.The model requires a input as texture , so here we used face crop texture to crop the face and pass as input to the ML component which then returns float value for every category. The value of each category is then converted to percentage and stored in text component.If no face is detected then the ML component stops and wait until any face is detected.

Challenges we ran into

The major challenge which I faced was to get the correct dataset for training the ML model.After lot of research I got SCUT-FBP5500-Database dataset which has 5500 images with their scores from 1 to 5. Further then I filtered the dataset and categorised into 5 category with order (good<average<beautiful<attractive<gorgeous) where good is the least one and gorgeous is the top most. The model took more then 8 hours to train with accuracy of 80%.

Accomplishments that we're proud of

The lens is one of its kind as of now which predicts the beauty using the model which is trained. It took me more than a month to properly make it. Going to the API docs of lens studio.