Convolutional features with Unsupervised Feature Learning for interest point detection and patch description

Abstract

Interest point detection and patch description is of great significance in computer vision applications. In this paper, we present a convolutional approach to interest point detection and patch description, which could detect and descript invariant 2D-features form images of different view conditions of an object or scene. In contrast to other convolution approaches that is trained to represent data or solve a classification task, our network could learn to not only describe but detect 2D-featrures of an image and the unsupervised feature learning is applied to ensure the algorithm with accuracy and distinctiveness. Also, we present a comparison evaluation on benchmark datasets of affine covariant features, yielding the algorithm of competitive results compared to the traditional approaches on detection and description.