BIG Thanks to PyTorch: Zero to GANs - the motivation for this blog post.

Imagine a game which can interact with the user without touch input or heft consoles. Kinect made it possible. But here we are gonna see the same objective getting full filled using a webcam.

Okay, I’m kind of hyping it overly :P

Glimpse of the output:

What I did

Before I build a model, I need a dataset using which my model can learn to recognize my hand gestures.

On searching the web, I found this publicly available dataset which solves my objective. It has the hand signs of almost all of the english alphabets.

Dataset link: Click Here

I utilized 3 classes to build my model, why not all? simply to speed up my training time.

• Using Pytorch I took a pretrained ResNet34 model and removed its last layer and added a softmax to find the probability of an image to the respective 3 classes.

I have two notebooks and one structured py script

Notebook 1: Click Here

Will have the step by step code snippets from processing the dataset to training the model

Notebook 2: Click Here

Utilizing the trained weights of the model and integration of the solution with opencv and game concept script.

Structured PY script: Click Here

Trained model’s weights: Click Here

The END!