How does CNN deal with rotation invariant pictures?

Question

I am trying to make a CNN model . Training the image . Want to know that When we apply kernel on image and take out the features of images. That features are rotation invariant or we have to apply some rotation invariant techniques? . Few person on stack overflow says that max pooling does rotational invariance , some person says that there is rotational invariant CNN architecture . Give me solid reason that how CNN deal with rotational invariant pictures ? Elaborate the answer .

In machine learning , we do some features extraction techniques like SIFT , SURF etc. and apply some algorithm on it, their features are scale and rotation invariant . How about in CNN ?

Answer 1

With a standard CNN, features are not rotation invariant, and the are not rotation equivariant. They are translation equivariant, but not rotation equivariant.

If you would like the model to be rotation invariance, then there are fancy methods to provide that, and simple methods.

The fancy method is to use a rotation-equivariant neural network. There are many research papers on this subject. See, e.g., Deep Rotation Equivariant Network by Li et al., Learning Steerable Filters for Rotation Equivariant CNNs by Weiler et al., and many others.

The simple method is to use data augmentation. During training time, randomly rotate the input image before feeding it to the neural network. You might need to train the neural network for more epochs.

Data augmentation is a lot easier to implement, and my impression is that it gives results that are close to what can be achieved with the fancy methods. Therefore, I suggest you try data augmentation first.