LSDNet: Trainable Modification of LSD Algorithm for Real-Time Line Segment Detection

As of today, the best accuracy in line segment detection (LSD) is achieved byalgorithms based on convolutional neural networks - CNNs. Unfortunately, thesemethods utilize deep, heavy networks and are slower than traditionalmodel-based detectors. In this paper we build an accurate yet fast CNN- baseddetector, LSDNet, by incorporating a lightweight CNN into a classical LSDdetector. Specifically, we replace the first step of the original LSD algorithm- construction of line segments heatmap and tangent field from raw imagegradients - with a lightweight CNN, which is able to calculate more complex andrich features. The second part of the LSD algorithm is used with only minormodifications. Compared with several modern line segment detectors on standardWireframe dataset, the proposed LSDNet provides the highest speed (amongCNN-based detectors) of 214 FPS with a competitive accuracy of 78 Fh . Althoughthe best-reported accuracy is 83 Fh at 33 FPS, we speculate that the observedaccuracy gap is caused by errors in annotations and the actual gap issignificantly lower. We point out systematic inconsistencies in the annotationsof popular line detection benchmarks - Wireframe and York Urban, carefullyreannotate a subset of images and show that (i) existing detectors haveimproved quality on updated annotations without retraining, suggesting that newannotations correlate better with the notion of correct line segment detection;(ii) the gap between accuracies of our detector and others diminishes tonegligible 0.2 Fh , with our method being the fastest.