Research map / Architectures and efficiency / RNNs and normalization
Driving by the Rules: A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map
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.
Introduces MapDR, a dataset of over 10,000 video clips with formatted lane-level traffic sign rules and rule-lane correspondences, plus a modular (VLE+MEE) and an end-to-end (RuleVLM) baseline for integrating traffic regulations into online HD maps.
Problem
Online HD map construction focuses on geometric and connectivity layers and neglects the traffic regulation layer, so traffic sign rules are not integrated into vectorized lane-level maps.
Approach
Decompose the task into rule extraction from traffic signs and rule-lane correspondence reasoning against vectorized centerlines, with a modular VLE+MEE pipeline and an MLLM (RuleVLM) that encodes centerlines via a map element encoder into Qwen-VL.
Key results
- VLE-MEE modular approach reaches overall F1 of 0.653 (Pall 63.35%, Rall 67.37%) on MapDR, far above the heuristic baseline at 0.035.
- RuleVLM achieves overall F1 0.642 and rule extraction P/R around 89%, outperforming Qwen-VL(TextPrompt) at 0.083 overall F1.
- Encoding centerlines as text in Qwen-VL performs poorly (overall F1 0.083), while visual prompting (0.392) and MEE-based encoding (0.642) improve substantially.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace heuristic rule extraction and lane matching with VLE-MEE | MapDR modular approach · MapDR overall | F1 score: 0.035 -> 0.653 | Helped |
| Use VLE-MEE instead of ALBEF-BERT | MapDR modular approach · MapDR rule-lane correspondence | Recall (%): 17.25 -> 82.16 | Helped |
| Use RuleVLM instead of Qwen-VL (TextPrompt) | Qwen-VL-Chat (9.6B) · MapDR overall | F1 score: 0.083 -> 0.642 | Helped |
| Encode centerlines with MEE rather than visual prompts | Qwen-VL-Chat (9.6B) · MapDR overall | F1 score: 0.392 -> 0.642 | Helped |
| Add intra- and inter-instance attention (ablation) | VLE · MapDR rule extraction | Recall (%): 57.56 -> 71.75 | Helped |
| Add text layout encoding (ablation) | VLE with attention · MapDR rule extraction | Recall (%): 71.75 -> 74.54 | Helped |
| Add text layout encoding (ablation) | VLE with attention · MapDR rule extraction | Precision (%): 76.86 -> 76.67 | Hurt |
| Add intra- and inter-instance attention (ablation) | MEE · MapDR rule-lane correspondence | Recall (%): 17.25 -> 71.39 | Helped |
| Add instance embedding (ablation) | MEE with attention · MapDR rule-lane correspondence | Recall (%): 71.39 -> 72.76 | Helped |
| Add type embedding (ablation) | MEE with attention and instance embedding · MapDR rule-lane correspondence | Recall (%): 72.76 -> 82.16 | Helped |
Surprising findings
- Textual representation of centerline coordinates degrades MLLM spatial reasoning and rule extraction, performing worse than visual prompting.
Limitations
- Zero-shot existing MLLMs struggled to produce stable structured outputs.
- Data limited to three Chinese cities (Beijing, Shanghai, Guangzhou).
Experimental setup
Datasets: MapDR
Benchmarks: MapDR
Metrics: precision recall F1
Baselines: Heuristic OCR and nearest-lane matching ALBEF-BERT Qwen-VL(TextPrompt) Qwen-VL(VisualPrompt)
Builds on: ALBEF Qwen-VL MapTRv2 OpenLane-V2
Closest papers by meaning
- Unveiling the Hidden: Online Vectorized HD Map Construction with Clip-Level Token Interaction and Propagation (NeurIPS 2024 · similarity 0.89)
- DriveCombo: Benchmarking Compositional Traffic Rule Reasoning in Autonomous Driving (CVPR 2026 · similarity 0.87)
- MapExpert: Online HD Map Construction with Simple and Efficient Sparse Map Element Expert (AAAI 2025 · similarity 0.87)
- The Road Ahead in Autonomous Driving: The KITScenes Multimodal Dataset (NeurIPS 2026 · similarity 0.86)
- LaneSegNet: Map Learning with Lane Segment Perception for Autonomous Driving (ICLR 2024 · similarity 0.86)
- Controllable Road Marking Generation (NeurIPS 2026 · similarity 0.85)
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.