Learning Spatial and Geometric Information for Robust Features

Abstract

Robust local feature detection and description are crucial in a huge of computer vision tasks. However, current feature descriptors lack geometric information and spatial features. In this paper, we propose a novel spatial and geometric feature fusion architecture, which can effectively tackle these problems. The proposed architecture includes three aspects. 1) A multiple low-level feature fusion (ML2F2) subnetworks, which could improve the ability to extract geometric information through the weighted fusion features. 2) A subnetwork for gradient feature extraction (GFE) which could validly extract the spatial features by encoding the horizontal and vertical gradients of an image. 3) Effective spatial geometric feature fusion (SGFF) module to alternatively fuse the above subnetworks. Experiments on the task of Image Matching and Long-Term Visual Localization show that the proposed method is superior to most advanced local feature descriptors.