HDD-Net: Hybrid Detector Descriptor with Mutual Interactive Learning

  • 2020-05-12 13:55:04
  • Axel Barroso-Laguna, Yannick Verdie, Benjamin Busam, Krystian Mikolajczyk
  • 50

Abstract

Local feature extraction remains an active research area due to the advancesin fields such as SLAM, 3D reconstructions, or AR applications. The success inthese applications relies on the performance of the feature detector anddescriptor. While the detector-descriptor interaction of most methods is basedon unifying in single network detections and descriptors, we propose a methodthat treats both extractions independently and focuses on their interaction inthe learning process rather than by parameter sharing. We formulate theclassical hard-mining triplet loss as a new detector optimisation term torefine candidate positions based on the descriptor map. We propose a densedescriptor that uses a multi-scale approach and a hybrid combination ofhand-crafted and learned features to obtain rotation and scale robustness bydesign. We evaluate our method extensively on different benchmarks and showimprovements over the state of the art in terms of image matching on HPatchesand 3D reconstruction quality while keeping on par on camera localisationtasks.

 

Quick Read (beta)

loading the full paper ...