GeoAT: Geometry-Aware Attention Feature Matching Network

We created GeoAT, an innovative step-by-step image feature matching model that leverages geometric information to match stepwise from coarse to exact. GeoAT is divided into three phases – coarse matching, mediate matching, and fine matching. It mainly uses self-attention and cross-attention mechanisms, combining features at different scales as well as positional information, so as to achieve a very precise match. One of the main innovations of GeoAT is the use of affine transformation matrices obtained from the coarse stage to guide the cross-attention process in the intermediate matching stage, which enables the neural network to concentrate primarily on areas containing possible correspondences while paying minimal attention to unrelated sections, consequently enhancing both the precision and efficiency of the correspondence identification process. GeoAT uses a flexible window-attention approach that can be adjusted to different scene conditions, allowing it to perform better when dealing with images with large differences in angles or less texture, and still get sufficiently accurate matching results at sufficiently high thresholds. We tested GeoAT on two well-known datasets, HPatches and MegaDepth. The test results show that GeoAT outperforms the current popular method in several important aspects.