Rotational equivariance in a convolutional neural network?

Question

I would like to know if the underlying CNN architecture has the property of rotational equivariance? I only know translational equivariance, but I'm not sure about rotation.

From my search, rotational equivariance can be achieved by rotating the input image for training. Do I need to do this? How big is the degree of rotation? To add more contex, for example, I have CNN that can detect / read text in landscape mode. If I rotate the image 90 degrees / make it a portrait, will it give the same result / do the same as the original?

Answer 1

You have scale and rotational invariance only to a certain extent - how accurately your setup can depend. You have this because pools containing functions potentially overlap.

What you offer is certainly possible. You can always change your training data by adding noise, rotation, various scales, etc., to achieve this goal. However, your model will still not be fully rotatable-invariant. It is also possible to change the network to achieve the “right” goal. I'm sure you stumbled upon Tiled CNN during your research (if not, you should definitely read this article). They use TICA for bickering, finding invariant features in the process.

90 °: . , , (, ), , 0 ° ( ) . .