Towards Light-Weight and Real-Time Line Segment Detection

Previous deep learning-based line segment detection (LSD) suffers from the immense model size and high computational cost for prediction. This constrains them real-time inference on computationally restricted environments. In this paper, we propose a light-weight detector resource-constrained environments named Mobile LSD (M-LSD). We design an extremely efficient architecture by minimizing backbone network removing typical multi-module process prediction found in previous methods. To maintain competitive performance with network, present novel training schemes: Segments of Line (SoL) augmentation, matching geometric loss. SoL augmentation splits into multiple subparts, which are used to provide auxiliary data during process. Moreover, loss allow capture additional cues. Compared TP-LSD-Lite, previously best method, our (M-LSD-tiny) achieves 2.5% increase 130.5% speed GPU. Furthermore, runs at 56.8 FPS 48.6 latest Android iPhone mobile devices, respectively. knowledge, is first available devices.