Learnable Line Segment Descriptor for Visual SLAM

Traditionally, the indirect visual motion estimation and simultaneous localization and mapping (SLAM) systems were based on point features. In recent years, several SLAM systems that use lines as primitives were suggested. Despite the extra robustness and accuracy brought by the line segment matching, the line segment descriptors used in such systems were hand-crafted, and therefore sub-optimal. In this paper, we suggest applying descriptor learning to construct line segment descriptors optimized for matching tasks. We show how such descriptors can be constructed on top of a deep yet lightweight fully-convolutional neural network. The coefficients of this network are trained using an automatically collected dataset of matching and non-matching line segments. The use of the fully-convolutional network ensures that the bulk of the computations needed to compute descriptors is shared among the multiple line segments in the same image, enabling efficient implementation. We show that the learned line segment descriptors outperform the previously suggested hand-crafted line segment descriptors both in isolation (i.e., for the subtask of distinguishing matching and non-matching line segments), but also when built into the SLAM system. We construct a new line based SLAM pipeline built upon a state-of-the-art point-only system. We demonstrate generalization of the learned parameters of the descriptor network between two well-known datasets for autonomous driving and indoor micro aerial vehicle navigation.