Comparative Study of Deep Learning Based Features in SLAM

Abstract

With the development of deep learning technology, plenty of learning-based feature detectors and descriptors methods are proposed. They achieve great performance on some specific datasets or environments. However, there are few experiments evaluating the performance of these learning-based features detectors and descriptors on a SLAM system. In this paper, we evaluate several learning-based feature detectors and descriptors on KITTI dataset with a SLAM system, using the accuracy of tracking as a metric of these methods. The results show that these learning-based feature detection methods perform poor on KITTI outdoor dataset.