Evaluation of Keypoint Extraction and Matching for Pose Estimation Using Pairs of Spherical Images

Keypoint extraction and matching has been widely studied by the computer vision community, mostly focused on pinhole camera models. In this paper we perform a comparative analysis of four keypoint extraction algorithms applied to full spherical images, particularly in the context of pose estimation. Two of the methods chosen for the comparative study, namely A-KAZE and ASIFT, have been designed considering a perspective camera model, but were already applied in an omnidirectional structure from motion pipeline, generating successful results in the literature. The other two algorithms are properly adapted versions of the traditional descriptors SIFT and ORB to the spherical domain, subbed SSFIT and SPHORB. We conduct our tests on captures of omnidirectional cameras, both synthetic and real, arbitrarily translated and rotated with known ground-truth transformations. The extracted keypoints are fed to the well-known 8-point algorithm with RANSAC, allowing to estimate the relative camera poses. These poses (translation vector and rotation matrix) are then compared to the ground-truth transformation parameters, generating the error metrics used in our analysis. Our results indicated that spherical descriptors SSIFT and SPHORB did not produce better results than planar descriptors A-KAZE and ASIFT in the context of pose estimation, particularly in the evaluation with real image pairs.