PaperlayerView exact revision on arXiv

Source-linked AI summary

Are Vision LLMs Road-Ready? A Comprehensive Benchmark for Safety-Critical Driving Video Understanding

Tong Zeng, Longfeng Wu, Liang Shi, Dawei Zhou, Feng Guo

arXiv:2504.14526v1cs.CVcs.CL

TL;DR

Safety-critical driving remains underexplored by VLLM benchmarks that inadequately test temporal-spatial video understanding. DVBench provides a hierarchical benchmark for safety-critical driving videos, and evaluations found substantial performance gaps, with no model surpassing 40% accuracy under GroupEval.

  • Problem

    Existing VLLM evaluations offer limited evidence for safety-critical driving, where rare incidents and static or text-focused benchmarks inadequately test temporal-spatial video perception and reasoning.

    Source: §1 · pp. 1–2

  • Method

    DVBench evaluates safety-critical driving-video perception and reasoning with a hierarchical taxonomy of 25 abilities and 10,000 curated multiple-choice questions.

    Source: §1 · p. 2, §3.1 · p. 3

  • Results

    Across 14 VLLMs, substantial performance gaps remained, with none surpassing 40% accuracy under GroupEval; fine-tuning yielded gains up to 10.94 percentage points.

    Source: §5 · p. 8

  • Takeaways & Limitations

    The results underscore targeted adaptation as necessary to bridge the gap between general-purpose VLLMs and mission-critical driving applications.

    Source: §5 · p. 8

    Quality control found unclear visual labels and skewed answer distributions, and excluded questions consistently answered correctly or incorrectly by all VLLMs.

    Source: §3.3 · p. 6

Abstract

from arXiv · show

Vision Large Language Models (VLLMs) have demonstrated impressive capabilities in general visual tasks such as image captioning and visual question answering. However, their effectiveness in specialized, safety-critical domains like autonomous driving remains largely unexplored. Autonomous driving systems require sophisticated scene understanding in complex environments, yet existing multimodal benchmarks primarily focus on normal driving conditions, failing to adequately assess VLLMs' performance in safety-critical scenarios. To address this, we introduce DVBench, a pioneering benchmark designed to evaluate the performance of VLLMs in understanding safety-critical driving videos. Built around a hierarchical ability taxonomy that aligns with widely adopted frameworks for describing driving scenarios used in assessing highly automated driving systems, DVBench features 10,000 multiple-choice questions with human-annotated ground-truth answers, enabling a comprehensive evaluation of VLLMs' capabilities in perception and reasoning. Experiments on 14 SOTA VLLMs, ranging from 0.5B to 72B parameters, reveal significant performance gaps, with no model achieving over 40% accuracy, highlighting critical limitations in understanding complex driving scenarios. To probe adaptability, we fine-tuned selected models using domain-specific data from DVBench, achieving accuracy gains ranging from 5.24 to 10.94 percentage points, with relative improvements of up to 43.59%. This improvement underscores the necessity of targeted adaptation to bridge the gap between general-purpose VLLMs and mission-critical driving applications. DVBench establishes an essential evaluation framework and research roadmap for developing VLLMs that meet the safety and robustness requirements for real-world autonomous systems. We released the benchmark toolbox and the fine-tuned model at: https://github.com/tong-zeng/DVBench.git.

1 Introduction

DVBench is introduced to evaluate VLLMs’ perception and reasoning in safety-critical driving videos, addressing gaps in existing benchmarks. In the reported evaluation, no model exceeded 40% accuracy under GroupEval.

1 Introduction

  • No model among 14 state-of-the-art VLLMs exceeded 40% accuracy under GroupEval, and fine-tuning selected models with DVBench data improved performance.The reported evaluation found substantial performance gaps in safety-critical driving perception and reasoning.Source: Figure 1 · p. 1, §1 · p. 2
  • Prior benchmarks underrepresent rare safety-critical events, miss temporal-spatial video reasoning, and narrow evaluation through static inputs, specialized tasks, or templated questions.Imbalanced datasets can reward simplistic answers, while limited temporal and spatial context weakens assessment of perception, reasoning, and prediction in dynamic scenarios.Source: §1 · pp. 1–2
  • DVBench contributes 10,000 curated multiple-choice questions spanning 25 driving abilities, organized in a three-level taxonomy aligned with PEGASUS and NHTSA.The questions are curated from safety-critical videos, including crashes and near-crash incidents, and assess perception and reasoning in dynamic environments.Source: §1 · p. 2
  • DVBench pairs video and text, uses automated question generation with human annotations, and introduces GroupEval’s grouped-capability assessment and randomized options to reduce position bias.The taxonomy comprises 2 level-one, 10 level-two, and 25 level-three abilities.Source: §1 · p. 2
  • The authors release the DVBench evaluation toolbox and fine-tuned models to support accessibility and reproducibility.The benchmark also establishes structured evaluation protocols and infrastructure for comparing VLLMs in safety-critical driving scenarios.Source: §1 · p. 2

2 Related Work

Prior work applies VLLMs to driving perception, reasoning, planning, and video interaction, but existing studies and benchmarks leave comprehensive evaluation of safety-critical driving unresolved.

VLLMs for Driving Scene Understanding

  • Driving VLLM work covers spatial reasoning [35], video responses, event descriptions [30], multi-sensor understanding, end-to-end interaction [40], perception and control [8], planning [24], and safety-event description [29], but lacks overall quantitative evaluation.Source: §2 · pp. 2–3

VLLM Evaluation Benchmarks

  • General benchmarks MMBench, MV-Bench [15], and VideoMME [5] use structured tasks such as multiple-choice questions for consistent scoring, but do not address autonomous-driving challenges.Source: §2 · p. 3
  • Driving benchmarks target interpretable reasoning, multimodal scene QA [27], or action and scenery with uncertain free-form scoring [25], yet remain incomplete for safety-critical evaluation.NuScenes-QA includes 34,000 scenes and 460,000 question-answer pairs; LingoQA answers average 17.2 words.Source: §2 · p. 3

3 The Proposed DVBench

DVBench combines a hierarchical taxonomy of driving perception and reasoning with human-annotated multiple-choice questions built from real-world driving videos. Its construction includes question curation and quality checks that address challenges in evaluating safety-critical scenarios.

3.1 Hierarchical Ability Taxonomy

  • The ability hierarchy defines perception and reasoning as two L1 abilities, expanded into 10 L2 abilities and 25 L3 assessment criteria.It draws on six-layer driving-scene frameworks and; perception covers physical road and traffic information, while reasoning interprets objects, environmental conditions, and digital data.Source: Figure 2 · p. 3, §3.1 · p. 3
  • The video data comes from SHRP 2 naturalistic driving research, whose annotated safety-critical events include crashes and near-crashes and cover 75 event, driver, and environmental variables.The project used a four-phase quality-assurance workflow spanning protocol development, annotator training, active data reduction with ongoing checks, and final verification.Source: §3.1 · pp. 3–4

3.2 Multiple-Choice Question Bank Construction

  • DVBench contains about 10,000 human-annotated multiple-choice questions spanning 25 granular abilities, built from 4,000 clips covering normal driving, crashes, and near-crashes.The clips follow a 5:4:1 ratio for those event types and were extracted from 2,000 safety-critical videos. Each question is tied to a single video and has 2–4 answer choices, a ground-truth answer, and domain knowledge.Source: §3.2 · p. 4
  • Question construction used LLM-assisted rewriting, annotation-derived answers, filtered candidate choices, and supplementary domain explanations to improve alignment and reduce option-length bias.ChatGPT o1-preview and Qwen2.5-72B generated alternatives for manual selection; questions and LLM-generated materials were checked manually, and candidate choices were selected to be close in length to the correct answer.Source: §3.2 · p. 5

3.3 Quality Control

  • Quality control combined manual review with majority voting by multiple VLLMs to flag potentially misleading, overly simplistic, or ambiguous questions.Review included iterative sampling, checking, and correction to catch hallucinations or inconsistencies in rewritten questions and domain knowledge.Source: §3.3 · pp. 5–6
  • Quality checks identified benchmark challenges including visually unclear labels, skewed answers, extreme difficulty, similar options, ambiguous terms, and mixed answer formats.Examples include unclear lane counts, highly prevalent “No” answers for vehicle rollovers, and confusing choices such as “Rear-end, striking” versus “Rear-end, struck.”Source: §3.3 · p. 6

4 Experiments

DVBench experiments show that current VLLMs remain limited on safety-critical driving tasks, while stricter evaluation and domain-specific adaptation reveal both performance gaps and opportunities for improvement.

4.1 Experimental Settings

  • VLLMs were evaluated zero-shot on DVBench using a shared prompt template to support fair comparison.The experiments focused on VLLMs capable of video understanding.Source: §4.1 · p. 6

4.2 Evaluation Strategy

  • GroupEval presents each question N times with the correct answer rotated and other options shuffled, passing only questions answered correctly in every trial.The evaluation addresses the approximately 25% top-1 accuracy possible by guessing among four choices and position biases such as favoring the first or last option [3] [19]; it can stop early after an incorrect prediction.Source: §4.2 · p. 6

4.3 IndividualEval vs GroupEval Results

  • GroupEval widened Qwen2-VL-72B’s Top-1 accuracy advantage over Qwen2-VL-7B from 5.9% to 18%, while most models lost around 10% to 20% accuracy.GroupEval requires correct predictions across repeated trials and became the default strategy for subsequent experiments.Source: §4.3 · p. 6, Table 2 · p. 7

4.4 Domain Knowledge Matters

  • Most models gained accuracy with domain knowledge, but PLLaVA-13B [38] declined slightly, suggesting the added context can be misinterpreted or increase task complexity.The authors add clarifying driving concepts after each question to help models understand domain-specific wording.Source: Table 3 · p. 7, §4.4 · p. 7

4.5 Main Results

  • No VLLM exceeded 40% overall accuracy; LLaVA-One-Vision-72B led perception at 39.4%, while MiniCPM-V led reasoning at 36.4%.Qwen2-VL-72B led Environmental Conditions at 76.7%, and LLaVA-One-Vision-72B led Risk & Hazard Assessment at 46.6%. MiniCPM-V (8B) outperformed LLaVA-One-Vision-72B in perception and performed comparably in reasoning; results also varied with architecture, training data, and optimization.Source: §4.5 · pp. 7–8, Table 4 · p. 8

4.6 Impact of Fine-Tuning on VLLMs

  • +10.94 percentage points for Qwen2-VL-7B and +5.24 percentage points for Qwen2-VL-2B after fine-tuning on DVBench data.The models were fine-tuned on 2,880 human-annotated question-answer pairs; Qwen2-VL-72B was excluded due to computational constraints.Source: Table 5 · p. 8, §4.6 · p. 8

5 Conclusion and Future Work

DVBench benchmarks safety-critical driving video understanding and reveals substantial limitations in current VLLMs, while domain-specific fine-tuning improves accuracy. Future work will broaden question types, enable public continuous benchmarking, and add rare and long-tail driving events.

  • No evaluated VLLM exceeded 40% accuracy under GroupEval; domain-specific fine-tuning improved accuracy by up to 10.94 percentage points.Source: §5 · p. 8
  • DVBench evaluates VLLMs across three hierarchical levels and 25 essential driving safety abilities, using a framework rooted in widely adopted driving-scenario taxonomies.Source: §5 · p. 8
  • Future work will add True/False and open-ended questions, create a public-leaderboard VLLM arena, and refine the dataset with more rare and long-tail driving events.Source: §5 · p. 8

A Details of Hierarchical Ability Taxonomy

DVBench’s hierarchical ability taxonomy organizes visual-linguistic understanding into perception and reasoning, providing a framework for evaluating how VLLMs interpret and synthesize complex visual-textual scenarios.

A Details of Hierarchical Ability Taxonomy

  • Perception covers object recognition, environmental conditions, and infrastructure, while reasoning covers inference, prediction, and causal analysis.These are the taxonomy’s two fundamental components.Source: A Details of Hierarchical Ability… · p. 10

A.1 Perception

The perception taxonomy organizes driving-scene understanding into five categories: environmental conditions, physical infrastructure, operational constraints, objects, and zones.

A.1 Perception

  • Environmental Conditions cover visibility-affecting light and weather, along with road-surface states such as snowy, dry, or wet.Source: §A.1 · p. 10
  • Physical Infrastructure describes road geometry, including alignment and grade, and lane configuration, including contiguous and through lanes.Source: §A.1 · p. 10
  • Operational Constraints include traffic flow and density, traffic-control signs and signals, and visual obstructions that limit the driver’s line of sight.Source: §A.1 · p. 10
  • Objects include static objects or animals and other vehicles, pedestrians, or road users involved in the depicted event.Source: §A.1 · p. 10
  • Zones describe geographic surroundings, such as urban, school, residential, or business settings, and intersections or junctions affecting vehicle movement.Source: §A.1 · p. 10

A.2 Reasoning

The reasoning taxonomy divides safety-critical driving-video understanding into five abilities, spanning event interpretation, vehicle behavior, spatial context, hazards, and responsibility.

Event Understanding (EU)

  • Event Understanding (EU) covers incident characteristics, precipitating triggers, and crash severity inferred from the video.It asks what kind of safety-critical event occurred, what initiated it, and how severe a crash was.Source: §A.2 · p. 10

Behavior & Maneuver Analysis (BMA)

  • Behavior & Maneuver Analysis (BMA) evaluates vehicle responses before, during, and after an event, including the quality of maneuver decisions.Examples include preincident and evasive maneuvers and post-event control.Source: §A.2 · p. 10

Spatial Reasoning (SR)

  • Spatial Reasoning (SR) assesses road alignment, grade, locality, lane position, and other entities’ locations relative to the subject vehicle.The taxonomy includes relationships with vehicles, pedestrians, animals, or objects.Source: §A.2 · p. 10

Risk & Hazard Assessment (RHA)

  • Risk & Hazard Assessment (RHA) identifies environmental and traffic dangers, as well as obstructions that may require evasive action.These hazards are evaluated for their implications for the subject vehicle’s safety.Source: §A.2 · p. 10

Causal & Responsibility (CR)

  • Causal & Responsibility (CR) determines which driver or non-motorist, if any, erred and explains why the event occurred.Source: §A.2 · p. 10

B Model Configurations

The section outlines the evaluated VLLMs’ architectural components and parameter scales, alongside the scope of their DVBench performance reporting.

B Model Configurations

  • Table 6 summarizes each evaluated VLLM’s language backbone, vision encoder, and parameter count.These details provide an overview of model architecture and scale.Source: B Model Configurations · p. 10, Table 6 · p. 11
  • Table 7 reports GroupEval performance without domain knowledge for L2 abilities and overall category performance for L1 abilities.Its L2 breakdown covers environmental conditions, physical infrastructure, operational constraints, objects, zones, event understanding, behavior and maneuver analysis, spatial reasoning, risk and hazard assessment, and causal and responsibility abilities.Source: Table 7 · p. 11

C Model Performance without Domain Knowledge · D Prompt Templates

The appendix reports VLLM performance on DVBench without domain knowledge and describes the zero-shot prompts used for evaluation.

C Model Performance without Domain Knowledge

D Prompt Templates

E Detailed Statistics of the MCQ Bank

The MCQ bank contains approximately 10,000 questions across 25 Level 3 abilities, with counts organized hierarchically to support targeted assessment of driving capabilities.

E Detailed Statistics of the MCQ Bank

Loading 2504.14526v1…