Analysis and Implementation of Rotation-invariant Neural Network Architectures for Feature Extraction

Vincent Ress, Markus Braendle, Norbert Haala

Abstract: When aligning images from different domains, such as thermal (IR) and human-visible light(RGB) images, classical feature extraction methods such as SIFT or SURF encounter severelimitations. While new techniques utilizing CNNs enable corresponding assignments, theirinherent non-equivariance to rotations restricts possible areas of applications. Within ourwork adaptions to the network architecture and the trainings pipeline to achieve a rotation-equivariant behaviour are discussed. The D2-Net, which is based on the VGG16 architectureand gains remarkable performance especially with regard to changes of the image domain,was used as reference. Through analyses on the HPatches dataset, significantly improvedequivariance properties were achieved for all adaptation types investigated.