Research map / Vision-language and multimodal / Multimodal LLMs
RoadBench: Benchmarking MLLMs on Fine-Grained Spatial Understanding and Reasoning under Urban Road Scenarios
Jun Zhang, Xin Zhang, Jie Feng, Long Chen, Junhui Wang, Zhicheng Liu, Depeng Jin, Yong Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune on RoadBench train set | Qwen3-VL-8B · Cross-View Lane Counting | F1: 0.266 -> 0.540 | Improved |
| fine-tune on RoadBench train set | Qwen3-VL-8B · FPV Road Type Classification | accuracy: 0.539 -> 0.830 | Improved |
| fine-tune on RoadBench train set | Qwen3-VL-8B · BEV Road Network Correction | Frechet distance: 0.294 -> 0.154 | Improved |
| fine-tune on RoadBench train set | Qwen3-VL-8B · BEV Lane Designation Recognition | accuracy: 0.372 -> 0.338 | Improved |
| compare MLLMs with domain-knowledge rule baseline on lane designation | MLLMs · BEV Lane Designation Recognition | accuracy: rule 0.587 vs GPT-5 0.579 | Hurt |
| scale parameters from GPT-5-Mini to GPT-5 | GPT-5 · BEV Lane Counting | F1: 0.355 -> 0.305 | Hurt |
| scale Qwen2.5-VL 32B to 72B | Qwen2.5-VL · FPV Lane Designation Recognition | accuracy: 0.310 -> 0.471 | Mixed |
| FPV instead of BEV viewpoint | MLLMs · lane counting and designation | accuracy: GPT-5 lane designation 0.579 -> 0.626 | Improved |
| cross-view instead of single-view input | MLLMs · lane counting | F1: GPT-5 FPV 0.498 -> cross-view 0.400 | Mixed |
| remove image reference-line prompt (keep text only) (ablation) | GLM-4.5V · BEV tasks | performance | Hurt |
| remove text reference-line prompt (keep visual only) (ablation) | GLM-4.5V · BEV tasks | performance | No clear effect |
| adverse lighting in FPV images (ablation) | GPT-5, Gemini-2.5-Pro, GLM-4.5V · FPV lane tasks | performance | Hurt |
Closest papers by meaning
- RoadSceneBench: A Lightweight Benchmark for Mid-Level Road Scene Understanding (CVPR 2026 · similarity 0.91)
- MAPLM: A Real-World Large-Scale Vision-Language Benchmark for Map and Traffic Scene Understanding (CVPR 2024 · similarity 0.89)
- MMSI-Bench: A Benchmark for Multi-Image Spatial Intelligence (ICLR 2026 · similarity 0.89)
- UrBench: A Comprehensive Benchmark for Evaluating Large Multimodal Models in Multi-View Urban Scenarios (AAAI 2025 · similarity 0.88)
- From Indoor to Open World: Revealing the Spatial Reasoning Gap in MLLMs (CVPR 2026 · similarity 0.88)
- GTR-Bench: Evaluating Geo-Temporal Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.88)
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.
