Descriptor Matching for a Discrete Spherical Image With a Convolutional Neural Network

In this paper, we propose a method of extracting feature descriptors from discrete spherical images using convolutional neural networks (CNNs). First, a captured full-view image is mapped to a discrete spherical image. Second, the features-from-accelerated-segment test algorithm is used to extract feature points in the discrete spherical image. Finally, an unsupervised CNN is used to obtain the descriptors of patches around each feature point. In the experiments, we compare these descriptors' performance to the closest existing state-of-the-art feature descriptors of discrete spherical images, spherical oriented FAST and rotated BRIEF (SPHORB), for image pairs having different camera rotation, noise levels, and general motions. The experimental results demonstrate that our proposed CNN-based discrete spherical image feature descriptors clearly outperform SPHORB both in accuracy and robustness.