Spherical back-projection-based feature detection and matching for panoramic 3D reconstruction

Panoramic 3D reconstruction is crucial in various fields; however, existing methods struggle with detecting distinguishable features in highly distorted regions of panoramic images. To address this problem, we propose a spherical back-projection-based approach for feature detection and matching, enhancing the robustness of panoramic 3D reconstruction in distorted areas. The panoramic image is mapped onto a unit sphere and partitioned into five groups of undistorted images using the geodesic grid. The speeded-up robust features algorithm is applied to detect features from each group of undistorted images, and the detected keypoints and descriptors are reprojected back to the original panoramic image for matching. Structure-from-motion and multiview stereo are then conducted based on the matched results. Analysis of an experimental comparison with the open-source software OpenMVG and other classical feature detection and description methods demonstrates that our proposed method significantly improves feature extraction, matching accuracy, texture recovery, and panoramic 3D reconstruction quality.