Dai Xili ; Gong Haigang ; Wu Shuai ; Yuan Xiaojun ; Ma Yi

Abstract

We present a one-stage Fully Convolutional Line Parsing network (F-Clip) thatdetects line segments from images. The proposed network is very simple andflexible with variations that gracefully trade off between speed and accuracyfor different applications. F-Clip detects line segments in an end-to-endfashion by predicting each line's center position, length, and angle. Wefurther customize the design of convolution kernels of our fully convolutionalnetwork to effectively exploit the statistical priors of the distribution ofline angles in real image datasets. We conduct extensive experiments and showthat our method achieves a significantly better trade-off between efficiencyand accuracy, resulting in a real-time line detector at up to 73 FPS on asingle GPU. Such inference speed makes our method readily applicable toreal-time tasks without compromising any accuracy of previous methods.Moreover, when equipped with a performance-improving backbone network, F-Clipis able to significantly outperform all state-of-the-art line detectors onaccuracy at a similar or even higher frame rate. In other word, under sameinference speed, F-Clip always achieving best accuracy compare with othermethods. Source code https://github.com/Delay-Xili/F-Clip.