Effect of Parameter Optimization on Classical and Learning-based Image Matching Methods

Deep learning-based image matching methods are improved significantly during\nthe recent years. Although these methods are reported to outperform the\nclassical techniques, the performance of the classical methods is not examined\nin detail. In this study, we compare classical and learning-based methods by\nemploying mutual nearest neighbor search with ratio test and optimizing the\nratio test threshold to achieve the best performance on two different\nperformance metrics. After a fair comparison, the experimental results on\nHPatches dataset reveal that the performance gap between classical and\nlearning-based methods is not that significant. Throughout the experiments, we\ndemonstrated that SuperGlue is the state-of-the-art technique for the image\nmatching problem on HPatches dataset. However, if a single parameter, namely\nratio test threshold, is carefully optimized, a well-known traditional method\nSIFT performs quite close to SuperGlue and even outperforms in terms of mean\nmatching accuracy (MMA) under 1 and 2 pixel thresholds. Moreover, a recent\napproach, DFM, which only uses pre-trained VGG features as descriptors and\nratio test, is shown to outperform most of the well-trained learning-based\nmethods. Therefore, we conclude that the parameters of any classical method\nshould be analyzed carefully before comparing against a learning-based\ntechnique.\n