Learning Attraction Field Representation for Robust Line Segment Detection

Abstract

This paper presents a region-partition based attraction field dual representation for line segment maps, and thus poses the problem of line segment detection (LSD) as the region coloring problem. The latter is then addressed by learning deep convolutional neural networks (ConvNets) for accuracy, robustness and efficiency. For a 2D line segment map, our dual representation consists of three components: (i) A region-partition map in which every pixel is assigned to one and only one line segment; (ii) An attraction field map in which every pixel in a partition region is encoded by its 2D projection vector w.r.t. the associated line segment; and (iii) A squeeze module which squashes the attraction field to a line segment map that almost perfectly recovers the input one. By leveraging the duality, we learn ConvNets to compute the attraction field maps for raw input images, followed by the squeeze module for LSD, in an endto-end manner. Our method rigorously addresses several challenges in LSD such as local ambiguity and class imbalance. Our method also harnesses the best practices developed in ConvNets based semantic segmentation methods such as the encoder-decoder architecture and the a-trous convolution. In experiments, our method is tested on the WireFrame dataset [1] and the YorkUrban dataset [2] with state-of-the-art performance obtained. Especially, we advance the performance by 4.5 percents on the WireFrame dataset. Our method is also fast with 6.6 ∼ 10.4 FPS, outperforming most of the existing line segment detectors.

Learning Regional Attraction for Line Segment Detection

“…This achievement is dramatically better than DWP [12], with a performance improvement of approximately 10 percent. Compared with the previous version AFM [33], AFM++ improves the F-measure by 5 percent on this dataset. This demonstrates the usefulness of outlier removal module and better optimizer, which will be further discussed below.…”

“…, 1.75 19 } for a-contrario validation where NFA is the number of false alarms. In addition, Linelet [18] uses the same thresholds as [18] 0.644 0.585 0.14 DWP [12] 0.728 0.627 2.24 AFM (U-Net) [33] 0.753 0.639 10.3 AFM (a-trous) [33] 0.774 0.647 6.6 AFM++ (a-trous) 0.823 0.672 8.0 [12] and YorkUrban dataset [4].…”

“…In the testing phase, the attraction field map computed by the network is squeezed to output line segments. Compared with the preliminary version of Attraction Field Map (AFM) [33], we further propose an outlier removal module based on the statistical priors of the training dataset, which significantly improves the performance of LSD. Besides, we find that better optimizer (e.g., the Adam optimizer [34]) with adaptive learning rate decay can make ConvNets learn better attraction field maps.…”

“…In this section, we evaluate the proposed line segment detector and compare with existing state-of-the-art line segment detectors [2], [3], [12], [18], [33] on the Wireframe dataset [12] and YorkUrban dataset [4]. The source code of this paper will be released at https://cherubicxn.github.io/ afmplusplus/.…”

“…3) The threshold of the line heat map is 10 for DWP; 4) The aspect ratio is set to 0.2 for AFM [33] and AFM++.…”