Research map / Detection and segmentation / CNN features
Semantic Line Combination Detector
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 SLCD, which generates all combinations of K reliable semantic lines and scores each combination jointly with semantic region grouping and compositional features, achieving state-of-the-art HIoU on SEL, NKL, and the new CDL dataset.
Problem
Semantic line detection methods score lines individually or by pairwise relations, so they fail to assess the overall harmony of a group of semantic lines that together describe the image composition.
Approach
Generate all 2^K combinations of K reliable line candidates and score each combination at once using semantic region grouping via cross-attention, a compositional feature (line, region, and positional maps), and a regressor trained with an HIoU-based score and a semantic region separation loss.
Key results
- SLCD reaches 84.09 HIoU on SEL, 3.06 points above the second-best HSLD (81.03).
- On CDL, SLCD achieves 68.85 HIoU versus 64.98 for HSLD, and on NKL 76.21 versus 74.29.
- On SEL_Hard, SLCD ranks second (68.15 HIoU) behind DRM (68.83) while running about 8.4 times faster than DRM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SLCD instead of HSLD | semantic line detection · SEL | HIoU: 81.03 -> 84.09 | Helped |
| use SLCD instead of DRM | semantic line detection · SEL_Hard | HIoU: 68.83 -> 68.15 | Hurt |
| use SLCD instead of HSLD | semantic line detection · NKL | HIoU: 74.29 -> 76.21 | Helped |
| use SLCD instead of HSLD | semantic line detection · CDL | HIoU: 64.98 -> 68.85 | Helped |
| set region query count to M=4, M=16, or M=32 instead of M=8 (ablation) | SLCD · CDL | HIoU: 68.85 -> 66.43; 68.85 -> 68.11; 68.85 -> 67.07 | Hurt |
| add the SRS loss (ablation) | SLCD · CDL | HIoU: 66.71 -> 68.85 | Helped |
| add region features to the line-feature-only variant (ablation) | SLCD compositional feature extraction · CDL | HIoU: 65.12 -> 66.37 | Helped |
| add the positional feature map to line and region features (ablation) | SLCD compositional feature extraction · CDL | HIoU: 66.37 -> 68.85 | Helped |
| increase reliable line count from K=6 to K=8 (ablation) | SLCD · CDL | HIoU: 67.96 -> 68.85 | Helped |
| increase reliable line count from K=8 to K=10 (ablation) | SLCD · CDL | HIoU: 68.85 -> 69.12 | Helped |
| increase line candidate count from N=400 to N=900 or N=1024 (ablation) | SLCD line detector · CDL | Recall: 88.61 -> 90.16; 88.61 -> 91.77 | Helped |
| increase line candidate count from N=1024 to N=1444 (ablation) | SLCD line detector · CDL | Recall: 91.77 -> 91.23 | Hurt |
Surprising findings
- DRM achieves higher HIoU on SEL_Hard but at roughly 8.4x the runtime of SLCD.
Limitations
- Exhaustive 2^K combination enumeration limits the number of reliable lines K (K=8 default) for runtime reasons.
- Performance on the Low and Front composition classes remains challenging due to complex object boundaries.
- AA at 1 degree for VP detection is lower than Zhou et al.
Experimental setup
Datasets: SEL NKL CDL ICCV NYU SYM_Hard AVA landscape CULane Oxford 5k Paris 6k ADE20K
Benchmarks: SEL SEL_Hard NKL CDL ICCV NYU SYM_Hard AVA landscape CULane
Metrics: HIoU AUC_P AUC_R AUC_F EA-score AA AUC_A spf
Baselines: SLNet DHT DRM HSLD Zhou et al. Jin et al.
Builds on: S-Net HSLD ResNet50
Closest papers by meaning
- ScaleLSD: Scalable Deep Line Segment Detection Streamlined (CVPR 2025 · similarity 0.82)
- Hierarchy-Agnostic Unsupervised Segmentation: Parsing Semantic Image Structure (NeurIPS 2024 · similarity 0.79)
- RankED: Addressing Imbalance and Uncertainty in Edge Detection Using Ranking-based Losses (CVPR 2024 · similarity 0.79)
- GCD: Advancing Vision-Language Models for Incremental Object Detection via Global Alignment and Correspondence Distillation (AAAI 2025 · similarity 0.79)
- CPN: Complementary Proposal Network for Unconstrained Text Detection (AAAI 2024 · similarity 0.79)
- Bi-SSC: Geometric-Semantic Bidirectional Fusion for Camera-based 3D Semantic Scene Completion (CVPR 2024 · similarity 0.79)
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.