Spatial Transformations in Deep Neural Networks

Convolutional Neural Networks (CNNs) have brought us the exceptionally significant improvement in the performance of the variety of visual tasks, such as object classification, semantic segmentation or linear regression. However, these powerful neural models suffer from the lack of spatial invariance. In this paper, we introduce the end-to-end system that is able to learn such invariance including in-plane and out-of-plane rotations. We performed extensive experiments on variations of widely known MNIST dataset, which consist of images subjected to deformations. Our comparative results show that we can successfully improve the classification score by implementing so-called Spatial Transformer module.