Research map / Detection and segmentation / CNN features
Improving Transformer Based Line Segment Detection with Matched Predicting and Re-ranking
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes RANK-LETR, a Transformer-based line segment detector that assigns each line to its centroid-nearest feature point (matched predicting, no bipartite matching), re-ranks predictions with learnable edge/endpoint/length cues, and adds a line segment ranking loss.
Problem
Transformer-based line segment detectors rely on bipartite matching, which causes slow convergence and need for long training, and their confidence scores often fail to reflect true segment quality, so accurate segments get suppressed.
Approach
Each feature point nearest the centroid of a line segment is responsible for predicting it, removing bipartite matching; a posterior re-ranking step rescores predictions using sampled endpoint, edge, and length scores from a learned geometric extractor; and a pairwise ranking loss ties confidence to prediction quality.
Key results
- On Wireframe, RANK-LETR reaches 65.0 sAP5 and 69.7 sAP10, outperforming the best listed baselines (HAWP 62.5/66.5, M-LSD† 63.3/67.1).
- Re-ranking with ground truth raises sAP5 from 62.0 to 91.3, indicating large headroom in ranking.
- Trained for 120 epochs versus 825 for LETR, while running at 9 FPS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RANK-LETR instead of M-LSD† | Line segment detection · Wireframe | sAP5: 63.3 -> 65.0 | Helped |
| Use RANK-LETR instead of M-LSD† | Line segment detection · Wireframe | sAP10: 67.1 -> 69.7 | Helped |
| Use RANK-LETR instead of HAWP | Line segment detection · YUD | sAP10: 28.5 -> 30.1 | Helped |
| Add matched predicting (ablation) | Foundational RANK-LETR model · Wireframe | sAP5: 50.3 -> 60.2 | Helped |
| Add matched predicting (ablation) | Foundational RANK-LETR model · Wireframe | sAP10: 58.5 -> 65.9 | Helped |
| Add rotation augmentation (ablation) | Matched-predicting model · Wireframe | sAP5: 60.2 -> 61.7 | Helped |
| Add rotation augmentation (ablation) | Matched-predicting model · Wireframe | sAP10: 65.9 -> 67.4 | Helped |
| Use prediction resolution 128 instead of 64 (ablation) | RANK-LETR with re-ranking and rotation augmentation · Wireframe | sAP5: 63.4 -> 65.0 | Helped |
| Use prediction resolution 128 instead of 64 (ablation) | RANK-LETR with re-ranking and rotation augmentation · Wireframe | sAP10: 68.3 -> 69.7 | Helped |
| Increase hidden dimension from 128 to 256 (ablation) | RANK-LETR · Wireframe | sAP5: 65.0 -> 65.1 | No clear effect |
| Reduce hidden dimension from 128 to 64 (ablation) | RANK-LETR · Wireframe | sAP5: 65.0 -> 62.9 | Hurt |
| Vary the referring-point count from 4 to 2 or 8 (ablation) | RANK-LETR · Wireframe | sAP5: 64.4; 64.6; 65.0 | No clear effect |
Surprising findings
- Re-ranking with ground-truth information raises sAP5 from 62.0 to 91.3, showing confidence scores often misrank accurate segments.
Limitations
- Each feature point predicts only one line, so two segments with very close centroids can only yield one prediction.
- Slower than CNN baselines such as HAWP and M-LSD (9 FPS).
Experimental setup
Datasets: Wireframe YorkUrban
Benchmarks: Wireframe YorkUrban
Metrics: sAP5 sAP10 sF10 sF15 LAP FPS
Baselines: LSD DWP AFM LGNN TP-LSD LETR L-CNN HAWP M-LSD
Builds on: LETR Deformable DETR L-CNN AFM HAWP
Closest papers by meaning
- RANK++LETR: Learn to Rank and Optimize Candidates for Line Segment Detection (NeurIPS 2025 · similarity 0.92)
- ScaleLSD: Scalable Deep Line Segment Detection Streamlined (CVPR 2025 · similarity 0.85)
- A Siamese Transformer with Hierarchical Refinement for Lane Detection (NeurIPS 2024 · similarity 0.84)
- ETO:Efficient Transformer-based Local Feature Matching by Organizing Multiple Homography Hypotheses (NeurIPS 2024 · similarity 0.83)
- TGBFormer: Transformer-GraphFormer Blender Network for Video Object Detection (AAAI 2025 · similarity 0.82)
- DECO: Unleashing the Potential of ConvNets for Query-based Detection and Segmentation (ICLR 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.