Joint Graph Learning and Matching for Semantic Feature Correspondence

In recent years, powered by the learned discriminative representation via graph neural network (GNN) models, deep matching methods have made great progresses in task of semantic features. However, these usually rely on heuristically generated patterns, which may introduce unreliable relationships to hurt performance. this paper, we propose a joint learning and network, named GLAM, explore reliable structures for boosting matching. GLAM adopts pure attention-based framework both Specifically, it employs two types attention mechanisms, self-attention cross-attention task. The discovers between features further update feature representations over learnt structures; computes cross-graph correlations sets be matched reconstruction. Moreover, final solution is directly derived from output layer, without employing specific decision module. proposed method evaluated three popular visual benchmarks (Pascal VOC, Willow Object SPair-71k), outperforms previous state-of-the-art all benchmarks. Furthermore, patterns our model are validated able remarkably enhance replacing their handcrafted with ones.