Deep Feature-based RGB-D Odometry using SuperPoint and SuperGlue
Satoshi Fujimoto, Nobutomo Matsunaga
Abstract
This paper presents a deep feature-based RGB-D odometry system using SuperPoint and SuperGlue. Geometry- based visual odometry systems face challenges, such as tracking failures in difficult scenes and scale ambiguity. As for the scale ambiguity problem, the map need not be initialized because 3D information can be obtained by using the depth cameras installed in smartphones in recent years. By contrast, learning- based visual odometry systems can estimate even particularly difficult scenes compared with ORB, SIFT, and LIFT features. We integrated this into our RGB-D odometry system. Compared with geometry-based and learning-based visual odometry systems, the proposed deep feature-based RGB-D odometry system achieved higher accuracy.