Research map / Vision-language and multimodal / Multimodal LLMs

RoadSceneBench: A Lightweight Benchmark for Mid-Level Road Scene Understanding

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Spatial reasoning in VLMs · 0 citations (OpenAlex)

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 RoadSceneBench, a 2,341-clip (11,705-image) benchmark for mid-level road semantics, and MapVLM, a Qwen2.5-VL-7B model trained with SFT followed by HRRP-T (hierarchical relational + temporal-consistency rewards under GRPO), which outperforms closed and open VLMs on it.

Problem

Existing road-scene benchmarks target low-level perception (detection, segmentation), leaving mid-level road semantics such as lane topology, lane-change feasibility, and ego-lane index, and their reasoning consistency across frames, largely unevaluated for VLMs.

Approach

A lightweight, self-collected, per-frame labeled multi-frame benchmark of mid-level road attributes with logical dependencies, plus an RL training scheme (HRRP-T) that combines a hierarchical frame-level reward (scene, relational, semantic layers) with a temporal reward (smoothness and transition plausibility) optimized with GRPO.

Key results

• MapVLM (SFT+HRRP-T) reaches overall precision 75.78% and recall 72.17% on RoadSceneBench, versus 60.61% / 52.70% for the best baseline Gemini-2.5-Pro.
• On Ego-lane Index, SFT+HRRP-T raises recall from 50.37% (SFT) to 84.67%, with precision rising from 69.34% to 75.44%.
• SFT alone already reaches overall 72.14% precision and 67.25% recall, above all zero-shot baselines.

Surprising findings

• Several closed-source and open-source VLMs score below 30% precision on ego-lane index and lane-change feasibility, despite strong performance on road scene classification.

Limitations

• Dataset is restricted to 20 Chinese cities due to policy constraints, limiting geographic diversity.
• Annotations are semi-automated, using pseudo-labels from in-house models reviewed by annotators, which may propagate their errors.
• Evaluation uses a single benchmark with no train/test split or variance reported; HRRP-T gains on road topology recall drop (42.73% vs 45.81% for SFT).

Experimental setup

Datasets: RoadSceneBench

Benchmarks: RoadSceneBench

Metrics: precision recall

Baselines: GPT-4o Gemini-2.5-Pro Claude-3.7-Sonnet InternVL3-78B Qwen3-VL-8B Qwen2.5-VL-72B LLaVA-Onevision-7B DeepSeek-VL2

Builds on: GRPO self-critical sequence training LoRA

Closest papers by meaning

• MAPLM: A Real-World Large-Scale Vision-Language Benchmark for Map and Traffic Scene Understanding (CVPR 2024 · similarity 0.89)
• DriveSpatial: A Benchmark for Spatiotemporal Intelligence in VLMs for Autonomous Driving (NeurIPS 2026 · similarity 0.89)
• Explicit Geometric Chain-of-Thought for Vision-Language-Action in Autonomous Driving (NeurIPS 2026 · similarity 0.89)
• DriveCombo: Benchmarking Compositional Traffic Rule Reasoning in Autonomous Driving (CVPR 2026 · similarity 0.88)
• SURDS: Benchmarking Spatial Understanding and Reasoning in Driving Scenarios with Vision Language Models (NeurIPS 2025 · similarity 0.88)
• VLScene: Vision-Language Guidance Distillation for Camera-Based 3D Semantic Scene Completion (AAAI 2025 · similarity 0.88)