Lattice-Point Mutually Guided Ground-to-Aerial Feature Matching for Urban Scene Images

Ground-to-aerial feature matching bridges information from cross-view images, which enables optimized urban applications, e.g., pixel-level geolocating and complete 3-D reconstruction. However, ground aerial images typically suffer drastic changes in viewpoint, scale, illumination, together with repetitive patterns. Thus, direct of local features between is particularly difficult because the low similarity descriptors high ambiguity true–false match discrimination. For this challenging task, we propose a novel lattice-point mutually guided (LPMG) method article. We specifically address two key issues: 1) reducing descriptor variance 2) enhancing discriminability. The former solved by recovering geometry appearance underlying image region through automatic view rectification on images. latter circumvented replacing conventional mismatch removal an LPMG strategy. In strategy, topology structure repeated façade elements (i.e., lattice), reliable point seeds, are first extracted rectified Then, seeds guide self-similar lattice tiles views to be precisely aligned, thereby estimating accurate transformation model tile correspondences. Finally, estimated powerfully supervises differentiation true false matches entire putative set. Extensive experiments conducted several datasets show that our can obtain considerable number nearly pure correct significantly outperforming those existing methods.