Towards Light-weight and Real-time Line Segment Detection

Previous deep learning-based line segment detection (LSD) suffers from theimmense model size and high computational cost for line prediction. Thisconstrains them from real-time inference on computationally restrictedenvironments. In this paper, we propose a real-time and light-weight linesegment detector for resource-constrained environments named Mobile LSD(M-LSD). We design an extremely efficient LSD architecture by minimizing thebackbone network and removing the typical multi-module process for lineprediction found in previous methods. To maintain competitive performance witha light-weight network, we present novel training schemes: Segments of Linesegment (SoL) augmentation, matching and geometric loss. SoL augmentationsplits a line segment into multiple subparts, which are used to provideauxiliary line data during the training process. Moreover, the matching andgeometric loss allow a model to capture additional geometric cues. Comparedwith TP-LSD-Lite, previously the best real-time LSD method, our model(M-LSD-tiny) achieves competitive performance with 2.5% of model size and anincrease of 130.5% in inference speed on GPU. Furthermore, our model runs at56.8 FPS and 48.6 FPS on the latest Android and iPhone mobile devices,respectively. To the best of our knowledge, this is the first real-time deepLSD available on mobile devices. Our code is available.