2017/11/17 by Dmytro Mishkin, Mishkin, Dmytro, Filip Radenović +3 · 3 citations
Computer Science · #Advanced Image and Video Retrieval Techniques #Advanced Vision and Imaging #Computer Vision and Pattern Recognition (cs.CV) #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #Image Retrieval and Classification Techniques #Neural and Evolutionary Computing (cs.NE)
paper · pdf · doi:10.48550/arxiv.1711.06704
openalex publication_date 2017/11/17 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
A method for learning local affine-covariant regions is presented. We show\nthat maximizing geometric repeatability does not lead to local regions, a.k.a\nfeatures,that are reliably matched and this necessitates descriptor-based\nlearning. We explore factors that influence such learning and registration: the\nloss function, descriptor type, geometric parametrization and the trade-off\nbetween matchability and geometric accuracy and propose a novel hard\nnegative-constant loss function for learning of affine regions. The affine\nshape estimator -- AffNet -- trained with the hard negative-constant loss\noutperforms the state-of-the-art in bag-of-words image retrieval and wide\nbaseline stereo. The proposed training process does not require precisely\ngeometrically aligned patches.The source codes and trained weights are\navailable at https://github.com/ducha-aiki/affnet\n