Eventually, MDGAT-matcher finds high-quality correspondences and at the same time handles non-matching points appropriately. We train our matcher using a novel gap loss guiding the network to learn a ...