HAT-Match: Graph Transformer with Hybrid Attention for Two-View Correspondence Pruning

Feature correspondence, particularly distinguishing inliers (true matches) from outliers (false matches), remains a core challenge in geometric computer vision. We present HAT-Match, a novel Graph Transformer framework with Hybrid Attention for two-view correspondence pruning. HAT-Match integrates three types of attention mechanismsself-attention for modeling pairwise dependencies, SE-based channel attention for emphasizing salient feature channels, and global global structure-aware for capturing structure-aware consistency. To model local geometric relationships, we first generate coarse clusters using a permutation-equivariant graph pooling and unpooling mechanism, which serves as an initial grouping of potentially consistent correspondences. Based on the resulting embeddings, we further construct local graphs using a DGCNN-style k-nearest neighbor (KNN) strategy, enabling the modeling of fine-grained local dependencies. These local graphs are then passed through a hybrid attention module that jointly encodes local and global contextual features. To refine correspondence confidence, we apply graph Laplacian-based attention over the global graph, enhancing discriminative feature propagation. The entire architecture is integrated into a progressive pruning framework that iteratively removes outliers and updates correspondence weights. Extensive experiments demonstrate that HAT-Match achieves state-of-the-art results across various challenging tasks, including relative pose estimation and visual localization, on both indoor and outdoor datasets.