Leveraging CNNs for Panoramic Image Matching Based on Improved Cube Projection Model

Three-dimensional (3D) scene reconstruction plays an important role in digital cities, virtual reality, and simultaneous localization mapping (SLAM). In contrast to perspective images, a single panoramic image can contain the complete information because of wide field view. The extraction matching feature points is critical difficult part 3D using images. We attempted solve this problem convolutional neural networks (CNNs). Compared with traditional algorithms, SuperPoint (SP) SuperGlue (SG) algorithms have advantages for handling images distortions. However, rich content leads significant disadvantage these regard time loss. To address problem, we introduce Improved Cube Projection Model: First, projected into split-frame overlap six directions. Second, SP SG are used process parallel matching. Finally, mapped back through coordinate inverse mapping. Experimental results multiple environments indicated that algorithm not only guarantee number extracted accuracy point but also significantly reduce computation compared other commonly algorithms.