bclassified

Classifying beards of the champions

Comment

bclassified

Classifying beards of the champions

Inspiration

It's Movember!

What is b(eing)classified?

Faces in images are detected by a forward pass through neural network. The beard part of the face is then passed through a pre-trained clustering algorithm, that categorizes the beards into different classes. These are then framed in different colors live, so that this technology can be deployed on a video stream.

How did we do it?

We used a pre-trained face-detection neural network from the Intel openVINO library to detect the faces. With this we created a small database of same-sized face images by passing through a video of the World Beard Championships. These images were then used to train a clustering model. We tried both MeanShift and k-Means. The clustering model can then predict the labels of new samples live. Here the former showed more promising results but the latter scaled better on more data and was faster for live inference. Before using Clustering algorithms it is wise to do a Principle Component Analysis which we also tried but which also was too slow for live usage.

What else can bclassified?

This approach can be ported to various use cases that are at least equally interesting and fun. Furthermore our model can be enhanced to not just do inference but to apply transformations on recognized faces and beards. We see two main customer types: people who want to get rid of beards and people who can't get enough of beards.

Built With

Share this project:

Updates