SalientPoint: A Self-Supervised Keypoint Detection Based on Saliency of Descriptor

This paper introduces a novel self-supervised frame-work designed to train keypoint detectors model across various descriptor models, named SalientPoint. The core of our method includes three modules: Homographic Augmentation, Absolute Saliency Calculation and Relative Saliency Calculation. The purpose of Homographic Augmentation is mainly to providing a large number of learnable candidate keypoint labels, ensuring smooth convergence of the model. The purpose of Absolute/Relative Saliency Calculation is to select a sufficient number of reliable labels from candidate keypoint labels for different descriptors. SalienPoint provides salient keypoints which are indirectly defined by the descriptors and avoids huge computation associated with dense descriptors. Experimental results demonstrate that SalientPoint can effectively improve image matching performance of various descriptors on two benchmarks. An ablation experiment validate that each component under this framework is indispensable.