Overview
As vision-language models (VLMs) become standard tools for driving-scene understanding, a critical question remains: can they reason defensively in the weakly structured traffic environments typical of the Global South? A new study accepted at the ECCV 2026 Workshop on Safe and Defensive Autonomous Driving (SDAD) tackles this gap head-on.
Introducing BD-HazardVLM
The authors present BD-HazardVLM, a diagnostic benchmark built from 500 Bangladesh road-scene images โ 375 drawn from RSUD20K and 125 from TFP-BD. Each image is annotated along multiple dimensions:
- Hazard presence
- Primary hazard category
- Risk timing
- Severity
- Defensive actions
- Visible evidence
- Hallucination and Western-road-bias traps
Models Evaluated
The benchmark probes three widely used VLMs under a shared zero-shot task definition and structured output schema:
- Qwen2.5-VL-7B
- InternVL3-8B
- MiniCPM-V-4.6
Key Findings
Coarse Recognition Looks Deceptively Good
Binary hazard-presence accuracy ranges from 88.0% to 89.2% โ essentially indistinguishable from the 89.0% majority-class baseline. In other words, headline accuracy numbers mask a model that may simply be predicting the dominant class.
Fine-Grained Reasoning Collapses
When the task demands genuine defensive-driving understanding, performance drops sharply:
| Task | Accuracy |
|---|---|
| Hazard category | Below 40% |
| Risk timing | 1.6% โ 12.4% |
| Exact defensive-action matching | Below 2% |
Severity and Recall Failures
InternVL3-8B achieves the highest severity accuracy at 61.2% โ yet records zero recall on the 58 high-severity cases, a dangerous blind spot for any safety-critical deployment.
Mode Collapse and Blind Spots
Qwen exhibits severe category mode collapse, while all three models show weak latent-hazard anticipation and near-zero self-reported trap recognition โ meaning they fail to flag the hallucination and Western-road-bias traps embedded in the benchmark.
Why It Matters
The paper's central conclusion is stark: for the evaluated models, coarse hazard recognition does not imply reliable fine-grained defensive-driving reasoning. This gap is especially consequential in regions like Bangladesh, where traffic patterns, road structure, and hazard distributions differ substantially from the Western datasets on which most VLMs are trained and evaluated.
The findings make a clear case for geographically diverse safety evaluation as a prerequisite for deploying VLMs in autonomous and assisted driving systems worldwide.
Publication Details
- arXiv: 2610.20627 [cs.CV]
- Submitted: 28 Sep 2026
- Venue: ECCV 2026 Workshop on Safe and Defensive Autonomous Driving (SDAD)
- Length: 16 pages, 1 figure, 5 tables
- Subjects: Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG); Robotics (cs.RO)
- Authors: Abdullah Al Maruf, Md. Sajedul Islam, Tanmoy Mridha, Irfan Hossain Bhuiyan
- DOI: 10.48550/arXiv.2610.20627
via ArXiv CV