2019/04/10 by Yurun Tian, Xin Yu, Tian, Yurun +9 · 5 citations
Computer Science · #Advanced Image and Video Retrieval Techniques #Computer Vision and Pattern Recognition (cs.CV) #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #Multimodal Machine Learning Applications
paper · pdf · doi:10.48550/arxiv.1904.05019
openalex publication_date 2019/04/10 · openalex created_date 2020/07/16 · openalex updated_date 2026/07/28
Despite the fact that Second Order Similarity (SOS) has been used with\nsignificant success in tasks such as graph matching and clustering, it has not\nbeen exploited for learning local descriptors. In this work, we explore the\npotential of SOS in the field of descriptor learning by building upon the\nintuition that a positive pair of matching points should exhibit similar\ndistances with respect to other points in the embedding space. Thus, we propose\na novel regularization term, named Second Order Similarity Regularization\n(SOSR), that follows this principle. By incorporating SOSR into training, our\nlearned descriptor achieves state-of-the-art performance on several challenging\nbenchmarks containing distinct tasks ranging from local patch retrieval to\nstructure from motion. Furthermore, by designing a von Mises-Fischer\ndistribution based evaluation method, we link the utilization of the descriptor\nspace to the matching performance, thus demonstrating the effectiveness of our\nproposed SOSR. Extensive experimental results, empirical evidence, and in-depth\nanalysis are provided, indicating that SOSR can significantly boost the\nmatching performance of the learned descriptor.\n