bd-arsa-road-safety-visual-audit
收藏资源简介:
BD-ARSA是一个用于孟加拉国道路安全视觉审计的多任务视觉语言数据集,遵循当地政府工程部(LGED)的审计方法。该数据集包含21,947条记录,每条记录将道路图像与结构化安全审计结果配对。数据来源于三个层次:expert_gold(343条)- 来自LGED审计报告的人工专家危险裁剪图像,每条仅包含单一危险;expert_silver(707条)- 道路场景照片与专家审计结果配对,经过标准化重写和人工审核;streetview(20,897条)- 使用VLM教师模型生成完整审计的道路场景图像。数据集采用12类LGED危险分类法,涵盖道路标线、路肩状况、路边严重性、行人设施、视线障碍、交通标志、速度管理、交叉口、路堤安全、公交站点、排水和防滑性等类别。每条记录包含图像、位置描述、道路类型、土地利用、关键安全问题列表、危险详细列表(含名称、类别、观察结果和严重性)、总体风险等级(低/中/高)以及附加建议等字段。数据集支持三个多任务学习任务:危险生成(所有记录)、总体风险评估(所有记录)和建议生成(21,635条记录)。数据按位置不相交原则划分为训练集(16,082条)、验证集(2,418条)和测试集(3,447条),风险类别存在不平衡(训练集中低风险仅占约1.4%)。数据集附带微调提示模板和类别权重文件,适用于视觉语言模型训练、道路安全审计自动化、多任务学习研究等应用场景。
BD-ARSA is a multi-task vision-language dataset for road safety visual auditing in Bangladesh, following the audit methodology of the Local Government Engineering Department (LGED). The dataset contains 21,947 records, each pairing a road image with structured safety audit results. Data sources include three tiers: expert_gold (343 records) - manually curated hazard-cropped images from LGED audit reports, each containing only a single hazard; expert_silver (707 records) - road scene photos paired with expert audit results, standardized and human-reviewed; and streetview (20,897 records) - road scene images with full audits generated using a VLM teacher model. The dataset employs a 12-class LGED hazard taxonomy, covering categories such as road markings, shoulder conditions, roadside severity, pedestrian facilities, sight obstructions, traffic signs, speed management, intersections, embankment safety, bus stops, drainage, and skid resistance. Each record includes fields such as image, location description, road type, land use, key safety issues list, detailed hazard list (with name, category, observation, and severity), overall risk level (low/medium/high), and additional recommendations. The dataset supports three multi-task learning tasks: hazard generation (all records), overall risk assessment (all records), and recommendation generation (21,635 records). Data is split into training (16,082 records), validation (2,418 records), and test (3,447 records) sets based on location-disjoint principles, with risk class imbalance (low risk accounts for only about 1.4% in the training set). The dataset comes with fine-tuning prompt templates and class weight files, suitable for applications like vision-language model training, automated road safety auditing, and multi-task learning research.
数据集概述:BD-ARSA: Road Safety Visual Audit Dataset
- 许可证: CC BY 4.0
- 语言: 英语 (en)、孟加拉语 (bn)
- 任务类别: 图像到文本 (image-to-text)、视觉问答 (visual-question-answering)
- 标签: 道路安全、孟加拉国、LGED、视觉语言模型、多任务
- 数据集大小: 10K < n < 100K (总计 21,947 条记录)
- 配置: 默认配置,包含训练集 (train)、验证集 (validation)、测试集 (test),数据以 Parquet 格式存储。
数据来源与结构
数据集由三个来源组成,按图像与危险标注的对应确定性分为三级(金 > 银 > 街景):
- expert_gold (343 条):来自 LGED 审计报告的人类专家危险裁剪图像,每张图像标注一个危险。覆盖 72 个审计地点。
- expert_silver (707 条):道路场景照片,配有人工审核并重写的 LGED 专家审计结果。
- streetview (20,897 条):街景图像,由 VLM 教师模型生成并验证的全审计标注。
数据模式(每条记录)
- record_id (字符串): 全局唯一标识符
- source (字符串): "expert_gold" | "expert_silver" | "streetview"
- district (字符串 | null)
- location_id (字符串)
- image (PIL.Image): 解码后的图像
- location_description (字符串 | null): 场景描述
- road_type (字符串 | null)
- land_use (字符串 | null)
- critical_safety_issues (列表[字符串]): 简短的“类别 - 问题”短语
- hazards (列表[{ hazard_name, hazard_category, observation, severity }]): 危险列表
- overall_risk_level (字符串): "Low" | "Medium" | "High"
- additional_recommendations (字符串 | null)
- tasks_available (子集): {"hazard_generation", "overall_risk", "recommendation"}
- split (字符串): "train" | "validation" | "test"
- source_expert_finding (字符串 | null): 仅 expert_silver 记录有,记录来源
hazard_category 为 12 个 LGED 标准类别之一:road_markings、shoulder_condition、roadside_severity、pedestrian_facilities、vision_obstruction、traffic_signs、speed_management、intersection、embankment_safety、bus_stoppage、drainage、skid_resistance。
粒度说明
- expert_gold: 单危险裁剪图像,
hazards只有一个条目(无 observation/severity),场景字段为 null。 - expert_silver 和 streetview: 全场景图像,多危险审计(平均每条约 5.5 和 4.2 个危险),包含 per-hazard 的 observation 和 severity。
多任务监督 (tasks_available)
- hazard_generation: 所有 21,947 条记录
- overall_risk: 所有 21,947 条记录
- recommendation: 21,635 条记录(全部 streetview + 全部 707 expert_silver + 31 expert_gold)
数据划分
- 划分原则: 按 location_id 划分,同一 location_id 不会出现在多个划分中。
- overall_risk_level 分布:
| 划分 | High | Medium | Low | 总计 |
|---|---|---|---|---|
| train | 9,517 | 6,340 | 225 | 16,082 |
| validation | 1,241 | 1,100 | 77 | 2,418 |
| test | 1,964 | 1,381 | 102 | 3,447 |
- 各来源在划分中的数量:
| 划分 | expert_gold | expert_silver | streetview | 总计 |
|---|---|---|---|---|
| train | 107 | 507 | 15,468 | 16,082 |
| validation | 78 | 96 | 2,244 | 2,418 |
| test | 158 | 104 | 3,185 | 3,447 |
类别不平衡
- Low 类别在训练集中占比约 1.4%。
- EG-ARSA 模型采用训练集特有的 logit 调整(τ=1)处理;同时提供训练集的逆频率权重:{ "Low": 23.83, "Medium": 0.85, "High": 0.56 }。
微调提示
- 数据集包含
prompts/finetune_prompts.py文件,提供无泄漏的微调提示和 JSON 目标序列化器。 prompts/generation_provenance.md文件记录了生成银标签的教师提示(用于可复现性,不建议直接用于微调)。
快速开始
python from datasets import load_dataset ds = load_dataset("Thamed-Chowdhury/bd-arsa-road-safety-visual-audit") sample = ds["train"][0] print(sample["record_id"], sample["overall_risk_level"]) print(sample["hazards"]) sample["image"] # PIL.Image
许可与引用
- 许可证: CC BY 4.0
- 引用格式: 参见 README 中的 BibTeX 条目(作者:Chowdhury, Md Thamed Bin Zaman 和 Hossain, Moazzem;年份:2026;代码仓库:https://github.com/Thamed-Chowdhury/EG-ARSA)
- 专家数据来源: 来自孟加拉国事故研究所(ARI, BUET)的现场道路安全审计,由世界银行资助的 RTIP-II 项目委托 LGED 执行。





