D2-Net: A Trainable CNN for Joint Detection and Description of Local\n Features

Abstract

In this work we address the problem of finding reliable pixel-level\ncorrespondences under difficult imaging conditions. We propose an approach\nwhere a single convolutional neural network plays a dual role: It is\nsimultaneously a dense feature descriptor and a feature detector. By postponing\nthe detection to a later stage, the obtained keypoints are more stable than\ntheir traditional counterparts based on early detection of low-level\nstructures. We show that this model can be trained using pixel correspondences\nextracted from readily available large-scale SfM reconstructions, without any\nfurther annotations. The proposed method obtains state-of-the-art performance\non both the difficult Aachen Day-Night localization dataset and the InLoc\nindoor localization benchmark, as well as competitive performance on other\nbenchmarks for image matching and 3D reconstruction.\n