Self-Supervised Bidirectional Learning for Graph Matching

Deep learning methods have demonstrated promising performance on the NP-hard Graph Matching (GM) problems. However, state-of-the-art usually require ground-truth labels, which may take extensive human efforts or be impractical to collect. In this paper, we present a robust self-supervised bidirectional method (IA-SSGM) tackle GM in an unsupervised manner. It involves affinity component and classic solver. Specifically, adopt Hungarian solver generate pseudo correspondence labels for simple probabilistic relaxation of matrix. addition, recycling consistency module is proposed samples by back permute input. imposes constraint between original one, theoretically supported help reduce matching error. Our further develops graph contrastive jointly with enhance its robustness against noise outliers real applications. Experiments deliver superior over previous state-of-the-arts five real-world benchmarks, especially under more difficult outlier scenarios, demon- strating effectiveness our method.