Abstract

We show how to train a Convolutional Neural Networkto assign a canonical orientation to feature points given animage patch centered on the feature point. Our method im-proves feature point matching upon the state-of-the art andcan be used in conjunction with any existing rotation sensi-tive descriptors. To avoid the tedious and almost impossi-ble task of finding a target orientation to learn, we proposeto use Siamese networks which implicitly find the optimalorientations during training. We also propose a new typeof activation function for Neural Networks that generalizesthe popular ReLU, maxout, and PReLU activation functions. This novel activation performs better for our task.We validate the effectiveness of our method extensively withfour existing datasets, including two non-planar datasets,as well as our own dataset. We show that we outperformthe state-of-the-art without the need of retraining for eachdataset.