Research map / Detection and segmentation / CNN features
Explicit Relational Reasoning Network for Scene Text Detection
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 ERRNet, a post-processing-free connected-component scene text detector that reformulates component grouping and ordering as an end-to-end tracking problem, plus a PIoU-based position-supervised classification loss.
Problem
CC-based scene text detection is accurate and well-suited to text shape but relies on slow, complex post-processing (grouping and ordering components), and classification confidence is misaligned with localization quality.
Approach
Treat each text instance as a sequence of ordered quadrilateral components moving over frames, so a tracking-style transformer decoder (intra-frame and inter-frame attention) with Hungarian sequence matching directly outputs grouped, ordered components; a Polygon Monte-Carlo PIoU estimate drives a position-supervised classification loss.
Key results
- ERRNet with pre-training reaches 89.4% F on CTW1500 and 90.3% F on MSRA-TD500, versus LRANet's 87.4% and 89.2%.
- Without pre-training, ERRNet achieves 87.2% F on CTW1500 vs TextBPN++ 86.5%, at 31.5 FPS on an RTX3090, about 16x faster than DRRG (2.0 FPS).
- The position-supervised classification loss adds 0.8% F on Total-Text for ERRNet and 0.7% F for DPText-DETR when applied to it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the ERR decoder (ablation) | ERRNet · CTW1500 | F-measure: 84.7 -> 87.2; 2.5% | Helped |
| Add the ERR decoder (ablation) | ERRNet · Total-Text | F-measure: 86.1 -> 88.1; 2.0% | Helped |
| Add position-supervised classification loss (PSC) (ablation) | ERRNet · Total-Text | F-measure: 87.3 -> 88.1; 0.8% | Helped |
| Add position-supervised classification loss (PSC) (ablation) | DPText-DETR · Total-Text | F-measure: 86.7 -> 87.4; 0.7% | Helped |
| Use explicit instead of implicit relational reasoning (ablation) | ERRNet · CTW1500 | F-measure: 86.3 -> 87.2 | Helped |
| Use explicit instead of implicit relational reasoning (ablation) | ERRNet · Total-Text | F-measure: 87.0 -> 88.1 | Helped |
| Increase input size from 608 to 704 (ablation) | ERRNet · CTW1500 | F-measure: 86.9 -> 87.2 | Helped |
| Increase input size from 800 to 1000 (ablation) | ERRNet · Total-Text | F-measure: 87.6 -> 88.1 | Helped |
| Increase component sequence length from 4 to 6 (ablation) | ERRNet · CTW1500 | F-measure: 86.8 -> 87.2; 0.4% | Helped |
| Increase component sequence length from 6 to 8 (ablation) | ERRNet · CTW1500 | F-measure: 87.2 -> 87.2 | No clear effect |
| Use B-spline instead of Bezier interpolation (ablation) | ERRNet TCI module · CTW1500 highly curved text subset | F-measure: 83.2 -> 84.3; 1.1% | Helped |
| Use PIoU instead of BIoU to measure localization quality (ablation) | ERRNet · CTW1500 | F-measure: 86.4 -> 87.2 | Helped |
Limitations
- Slightly slower than LRANet (31.5 vs 37.2 FPS).
- Input size trades accuracy for speed; performance drops at smaller input sizes (e.g., 85.2% F on Total-Text at 608).
Experimental setup
Datasets: Total-Text CTW1500 ArT MSRA-TD500 Synth150K MLT
Benchmarks: Total-Text CTW1500 ArT MSRA-TD500
Metrics: precision recall F-measure FPS
Baselines: DB TextBPN++ LRANet DPText-DETR DRRG ReLaText TPSNet CT-Net
Builds on: DPText-DETR Deformable DETR DRRG ReLaText
Closest papers by meaning
- EAST: An Efficient and Accurate Scene Text Detector (CVPR 2017 · similarity 0.86)
- Deep Relational Reasoning Graph Network for Arbitrary Shape Text Detection (CVPR 2020 · similarity 0.86)
- LRANet: Towards Accurate and Efficient Scene Text Detection with Low-Rank Approximation Network (AAAI 2024 · similarity 0.85)
- Arbitrary Reading Order Scene Text Spotter with Local Semantics Guidance (AAAI 2025 · similarity 0.84)
- CPN: Complementary Proposal Network for Unconstrained Text Detection (AAAI 2024 · similarity 0.84)
- Scene Graph Thinking: Reinforcing Structured Visual Reasoning for Multimodal Large Language Models (ICML 2026 · similarity 0.83)
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.