HDD-Net: Hybrid Detector Descriptor with Mutual Interactive Learning

Local feature extraction remains an active research area due to the advances in fields such as SLAM, 3D reconstructions, or AR applications. The success these applications relies on performance of detector, descriptor, and its matching process. While trend detector-descriptor interaction most methods is based unifying two into a single network, we propose alternative approach that treats both components independently focuses their during learning We formulate classical hard-mining triplet loss new detector optimisation term improve keypoint positions descriptor map. Moreover, introduce dense uses multi-scale within architecture hybrid combination hand-crafted learnt features obtain rotation scale robustness by design. evaluate our method extensively several benchmarks show improvements over state art terms image reconstruction quality while keeping par camera localisation tasks.