Skip to the content.

Spatial Transformer Networks

Contact me

本系列博客主页及相关见此处


Abstract

The ability of spatially invariant of CNN is limited. This paper introduces a Spatial Transformer module, which manipulate the data in spatial way, and can be insserted to any CNN network, without andy extra training. It achieved state-of-the-art performance.

1 Introduction

Spatial transformer can be helpful for mutifariout tasks:

3 Spatial Transformers

Spatial transformers mechanism is split into three parts:

3.1 Localisation Network

$\theta$ is varied on different transformation type, affine transformation $\theta$ is 6-dim. $\theta = f_{loc}(U)$

3.2 Parameterised Sampling Grid

only 6 params are reuqired to define $A_\theta$

Differentiable Image Sampling

Integer sampling kernel reduces to: