A2B: Anchor to Barycentric Coordinate for Robust Correspondence

There is a long-standing problem of repeated patterns in correspondence problems, where mismatches frequently occur because inherent ambiguity. The unique position information associated with makes coordinate representations useful supplement to appearance for improving feature correspondences. However, the issue appropriate representation has remained unresolved. In this study, we demonstrate that geometric-invariant representations, such as barycentric coordinates, can significantly reduce between features. first step establish theoretical foundation geometrically invariant coordinates. We present seed matching and filtering network (SMFNet) combines consistency coarse-to-fine strategy order acquire reliable sparse then introduce Degree, novel anchor-to-barycentric (A2B) encoding approach, which generates multiple affine-invariant coordinates from paired images. Degree be used plug-in standard descriptors, matchers, filters improve quality. Extensive experiments synthesized indoor outdoor datasets alleviates helps achieve state-of-the-art performance. Furthermore, also reports competitive performance third Image Matching Challenge at CVPR 2021. This approach offers new perspective alleviate emphasizes importance choosing