TransMatch: Transformer-based correspondence pruning via local and global consensus

Correspondence pruning aims to filter out false correspondences (a.k.a. outliers) from the initial feature correspondence set, which is pivotal to matching-based vision tasks, such as image registration. To solve this problem, most existing learning-based methods typically use a multilayer perceptron framework and several well-designed modules to capture local and global contexts. However, few studies have explored how local and global consensuses interact to form cohesive feature representations. This paper proposes a novel framework called TransMatch, which leverages the full power of Transformer structure to extract richer features and facilitate progressive local and global consensus learning. In addition to enhancing feature learning, Transformer is used as a powerful tool to connect the above two consensuses. Benefiting from Transformer, our TransMatch is surprisingly effective for differentiating correspondences. Experimental results on correspondence pruning and camera pose estimation demonstrate that the proposed TransMatch outperforms other state-of-the-art methods by a large margin. The code will be available at https://github.com/lyz8023lyp/TransMatch/.