chutya_rating
Uses CV and ML to rate you on a scale of - Kati Chutya lagte rahe ho -> to -> Arey boss?
Theory
Well, as you might have noticed this, project is for fun - like most of my projects, but there are quite buzz words involved in making it. I will use OpenCV, Numpy and scikit-learn to develop a completely automated pipeline that takes a photograph of a person’s face, and rates the photo on how chutya you are.
While applying ML Algo's to a computer vision problem, the high dimensionality of visual data presents a huge challenge. Even a relatively low-res 200×200 image translates to a feature vector of length 40,000. Which is a lot for my tiny machine to work with. Deep learning models like CNN's can work directly with raw images, but they require huge amounts of data to be successful which is agian something I do not have (unless someone wants to make a huge DB of chutiya's :P)
Feature Extraction
Anyways --> We will need to do some feature extraction instead of feeding the images directly to the machine.
This is we will we taking out the facial landmarks. The following image should clear out any doubt you have on facial landmarks
We will use openface to get the facial landmarks.
Dataset
Now that we have a way to extract the features, we require a dataset of images. Which is why i am writing this read me way before publishing the results
I need your help in creating a dataset, all I need are a few names of people who you think "SHAKAL SEY CHUTYA LAGTAE HAI"
^^ Old
Create your own dataset of humans and place then in pics folder labelled 1 - n.jpg
ML bit
Now this is where the things get interestin. We will be using scikit-learn for this.
Ubuntu
sudo apt-get install python-sklearn
Mac OSX
pip install -U numpy scipy scikit-learn
Windows
pip install -U scikit-learn
Once that is done --> in short, what we want to do here is calculate ratios between all possible pairs of points from our facial landmarks and then use that to train our machine.
To train our model, we will be using PCA -> since our feature set is large (hanged my text editor for a simple search) Using PCA in scikit-learn is pretty simple, and the following is all you need:
pca = decomposition.PCA(n_components=20)
pca.fit(features_train)
features_train = pca.transform(features_train)
features_test = pca.transform(features_test)
In the above code, n_components is a hyper-parameter, and it’s value can be chosen by testing the performance of the model on a validation set (or cross-validation).
I will be using linear regression to train my model you are free to use SVM, Random forest anything. To create a linear regression model you can use the following code
regr = linear_model.LinearRegression()
regr.fit(features_train, ratings_train)
ratings_predict = regr.predict(features_test)
The above code will fit a linear regressor on the traininng data, and then store the predictions.