Source-linked AI summary
Learning Attraction Field Representation for Robust Line Segment Detection
Nan Xue, Song Bai, Fudong Wang, Gui-Song Xia, Tianfu Wu, Liangpei Zhang
TL;DR
Line segment detection is difficult because existing approaches face local ambiguity, class imbalance, and heuristic post-processing. This paper recasts it as region coloring using a dual attraction-field representation learned by ConvNets, then reports state-of-the-art results on two benchmarks, including a 4.5% WireFrame improvement and 6.6 ∼10.4 FPS speed.
Problem
Existing line segment detectors face local ambiguity, class imbalance, and reliance on extra heuristics when inferring segments from line heat maps.
Method
The method learns ConvNets to predict region-partition-based attraction field maps from images and applies a squeeze module to recover line segments.
Results
State-of-the-art performance is reported on WireFrame and YorkUrban, with a 4.5% WireFrame improvement and 6.6 ∼10.4 FPS speed.
Takeaways & Limitations
The dual representation provides a direct region-coloring formulation that addresses local ambiguity and class imbalance while leveraging semantic-segmentation ConvNets.
Takeaways & Limitations
Intermediate attraction-field points can differ from annotated pixel locations when line segments are not strictly horizontal or vertical.
Abstract
from arXiv · showhide
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 in-put images, followed by the squeeze module for LSD, in an end-to-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 and the YorkUrban dataset 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 existing line segment detectors.
1. Introduction
The paper reformulates line segment detection as region coloring through a dual attraction-field representation, enabling end-to-end ConvNet prediction and squeeze-based reconstruction. This addresses local ambiguity and class imbalance while retaining efficient line-segment recovery.
- Motivation and Objective: Existing line segment detectors struggle with local ambiguity, class imbalance, multi-scale discretization, and heuristic post-processing.Two-stage methods generate line heat maps and then fit line models, often requiring additional heuristics or contextual supervision.
- Motivation and Objective: The proposed duality converts line segment detection into region coloring, allowing semantic-segmentation ConvNet practices to be applied.The formulation is designed to address local ambiguity and class imbalance in a principled way.
- Method Overview: The representation assigns every pixel to one line segment, encodes its projection vector in an attraction field, and uses a squeeze module to recover line segments.The partition regions are mutually exclusive and cover the image lattice; the squeeze module almost perfectly recovers the input line segment map.
- Method Overview: ConvNets are trained end-to-end to predict attraction field maps from raw images, after which the squeeze module produces line segments.The experiments use U-Net and a modified DeepLab V3+ network.
- Method Overview: 4.5% improvement is reported on the WireFrame dataset, with state-of-the-art performance on WireFrame and YorkUrban and 6.6 ∼10.4 FPS speed.The method is compared with existing line segment detectors on both benchmarks.
2. Related Work and Our Contributions
Prior line-segment methods rely on local handcrafted features or complex edge-junction pipelines. The paper introduces a dual region-attraction representation that directly supports deep learning and reports state-of-the-art benchmark performance.
- Related Work: Handcrafted local features are sensitive to imaging conditions and lack global context, limiting robust line-segment extraction.Noise, illumination, and locally similar textures can affect detection.
- Deep Edge and Line Segment Detection: Deep edge and line-segment methods improve representation learning, but existing pipelines still combine edge and junction detection before producing segments.The edge-junction relationship remains difficult to model directly.
- Our Contributions: The proposed dual representation bridges line segment maps and region-partition-based attraction field maps.The paper presents this as a novel representation for line segment detection.
- Our Contributions: Reformulating detection as region coloring opens access to semantic-segmentation methods for handling local ambiguity and class imbalance.The contribution is framed as a principled alternative to existing LSD formulations.
- Our Contributions: 4.5% improvement is reported on WireFrame, with state-of-the-art performance also obtained on YorkUrban.These are the paper's reported benchmark outcomes.
3. The Attraction Field Representation
The paper represents line segment maps through mutually exclusive support regions and per-pixel attraction vectors, then reverses this representation with a squeeze module that greedily groups vectors into segments. This dual formulation supports accurate recovery and reframes line segment detection as learning attraction fields for region coloring.
- 3.1. The Region-Partition Map: Each pixel is assigned to exactly one line segment, forming mutually exclusive regions whose union covers the image lattice.Regions are defined by comparing each pixel’s point-to-line-segment distance across all segments.
- 3.2. Computing the Attraction Field Map: An attraction field encodes every region pixel by the 2D vector from that pixel to its projection point on the associated segment.The projection uses the segment interior when applicable and the nearest endpoint otherwise.
- 3.3. The Squeeze Module: The squeeze module converts attraction vectors into line segments by discretizing projection points, collecting vectors into sparse line proposals, and greedily grouping aligned vectors.The grouping searches a local window using an angular threshold, grows candidate sets, and fits thin line segments from their support points.
- 3.3. The Squeeze Module: Candidate segments are retained only when their fitted support-region rectangle is sufficiently thin, while failed candidates release their support pixels for later searches.This verification step can deactivate the initiating pixel and reactivate the associated support set.
- 3.3. The Squeeze Module: On WireFrame, the representation’s average precision and recall exceed 0.99 and 0.93 across scales from 0.5 to 2.0, respectively.These results are reported as verification of the duality between attraction fields and line segment maps.
- 3. The Attraction Field Representation: The formulation turns line segment detection into region coloring, reducing local ambiguity and avoiding the line-versus-nonline class imbalance of gradient-based heat-map prediction.The learning system predicts attraction fields from input images and applies the squeeze module for detection.
4. Robust Line Segment Detector
The detector learns attraction field maps from images with encoder-decoder ConvNets, then applies a squeeze module to recover line segment maps. It uses U-Net and a-trous Residual U-Net designs with normalized and value-stretched training targets.
- Data Processing: Training constructs a dual dataset of images and attraction field maps, normalizes map values by image size, and applies an invertible stretching transformation.The transformation addresses numerically small normalized values during training and uses ε = 1e−6 to avoid log(0).
- Inference: The robust LSD predicts attraction field maps with a ConvNet and converts them into line segment maps through the squeeze module.The predicted map is the size-normalized and value-stretched attraction field map.
- Network Architectures: The method realizes attraction field prediction with U-Net and a-trous Residual U-Net architectures.The latter combines an ASPP module with residual skip connections.
- Network Architectures: The architectures use encoder-decoder stages with upsampling and skip-connected feature combination for attraction field learning.Table 1 specifies the investigated network configurations, including double-convolution and residual-block operators.
- Training: The networks are trained with L1 loss and standard stochastic-gradient optimization after augmentation and resizing images to 320 × 320.Training uses momentum 0.9, an initial learning rate of 0.01, 200 epochs, and decay by 0.1 every 50 epochs.
5. Experiments
The proposed detector is evaluated against established line segment methods on WireFrame and YorkUrban using precision, recall, F-measure, speed, and qualitative visualizations. It achieves the strongest reported overall performance while remaining faster than most competing detectors.
- Precision & Recall: The proposed method outperforms existing approaches on both WireFrame and YorkUrban, with deeper networks and ASPP further improving F-measure.YorkUrban scores are reduced for all methods because some line segments are unlabeled in this Manhattan-frame dataset.
- Visualization: Qualitative comparisons cover WireFrame, YorkUrban, and Internet images across LSD, MCMLSD, Linelet, Deep Wireframe Parser, the proposed model, and ground-truth or input references.The Internet images have no ground-truth annotations and are shown with input images as references.
- Visualization: The proposed method produces more complete line segments than previous approaches while avoiding junction-driven orientation errors and false detections.Deep Wireframe Parser also detects relatively complete segments, but junction misconnections and inaccurate branch orientations affect its detections.
- Visualization: Deep learning methods use global information to complete low-contrast line segments and suppress false detections in edge-like textures.LSD, MCMLSD, and Linelet remain limited by local features, although LSD has slightly better overall F-measure than Linelet and Linelet produces cleaner visualizations.
6. Conclusion
The paper builds a dual attraction-field representation that reformulates line segment detection as region coloring with learned ConvNets. Evaluations on two benchmarks report state-of-the-art performance and 6.6 ∼10.4 FPS speed.
- 6. Conclusion: The proposed dual representation reformulates line segment detection as a region-coloring problem addressed by learned convolutional neural networks.The formulation is designed to address local ambiguity and class imbalance while using semantic-segmentation practices.
- 6. Conclusion: The method combines a region-partition map, attraction field representation, and squeeze module to connect region-based predictions with line segment maps.The squeeze module converts the attraction field into line segments, while the broader framework uses encoder-decoder and a-trous convolution designs.