Barroso-Laguna Axel ; Riba Edgar ; Ponsa Daniel ; Mikolajczyk Krystian

Abstract

We introduce a novel approach for keypoint detection task that combineshandcrafted and learned CNN filters within a shallow multi-scale architecture.Handcrafted filters provide anchor structures for learned filters, whichlocalize, score and rank repeatable features. Scale-space representation isused within the network to extract keypoints at different levels. We design aloss function to detect robust features that exist across a range of scales andto maximize the repeatability score. Our Key.Net model is trained on datasynthetically created from ImageNet and evaluated on HPatches benchmark.Results show that our approach outperforms state-of-the-art detectors in termsof repeatability, matching performance and complexity.