LTM: efficient learning with triangular topology constraint for feature matching with heavy outliers

Image feature matching, which aims to establish correspondence between two images, is an important task in computer vision. Among image feature matching, the removal of mismatches is crucial to ensure the correctness of the matches. In recent years, machine learning has become a new perspective for mismatch removal. However, existing learning-based methods require a large amount of image data for training, which shows a lack of generalizability and is hard to deal with cases with high mismatch ratio. In this paper, we induce the triangular topology constraint into machine learning, where topology constraints around the matching points are summarized; combining with the idea of sampling, we achieve the task of removing mismatches. Topology constraints are studied in spite of the image input; our LTM (learning topology for matching) just needs fewer than 20 parameters as input, so that only ten training image pairs from four image sets involving about 3,000 matches are employed to train; it still achieves promising results on various datasets with different machine learning approaches. The experimental results of this study also demonstrate the superior performance of our LTM over existing methods.