驾驶场景识别数据集(DSI)
收藏资源简介:
驾驶场景识别数据集(DSI)是一个包含7个子集的单一标签数据集,每个子集根据一个特定的场景属性进行标注。该数据集共包含31835张场景图片,涵盖了7种不同的场景属性,如分离式基础设施、道路功能、天气、施工区域、天气相关道路状况、一天中的时间以及交叉路口类型。数据集的构建旨在为多标签驾驶场景分类提供全面的标注数据,从而解决现有数据集中标签不平衡和类分布不均的问题。DSI数据集的创建过程涉及从多个公共驾驶视频数据集中抽取样本,并补充了来自YouTube视频的额外图片。该数据集旨在支持自动驾驶车辆更好地理解、推理和与复杂的驾驶环境交互,以提高安全性。数据集的访问地址为https://github.com/KELISBU/KAA-CAL。
The Driving Scene Identification Dataset (DSI) is a single-label dataset consisting of 7 subsets, each annotated based on a specific scene attribute. It contains a total of 31,835 scene images, covering 7 distinct scene attributes including separated infrastructure, road function, weather, construction zones, weather-related road conditions, time of day, and intersection types. The dataset is constructed to provide comprehensive annotated data for multi-label driving scene classification, aiming to address the issues of label imbalance and uneven class distribution prevalent in existing datasets. The development of DSI involved extracting samples from multiple public driving video datasets, with additional images sourced from YouTube videos. This dataset is intended to support autonomous vehicles in better understanding, reasoning, and interacting with complex driving environments to enhance driving safety. The dataset is available at https://github.com/KELISBU/KAA-CAL.
KAA-CAL 数据集概述
数据集简介
- KAA:通过单任务学习从多个单标签数据集中获取和积累场景识别知识。
- CAL:有效解决单标签和多标签数据之间的知识差距问题。
Driving Scene Identification (DSI) 数据集
- 数据来源:从公开驾驶视频数据集(BDD100K、HSD、ROADWork Data)和YouTube视频中采样的场景图像。
- 数据规模:共31,835张场景图像。
- 标注类别:24个独立类别,涵盖7个场景属性。
场景属性及类别分布
-
Grade-separated Infrastructure
- 类别:Overhead bridges, Tunnels, Open roads
- 训练集:4,874;验证集:1,866;测试集:1,025
-
Road Function
- 类别:Local, Arterial, Interstate, Collector
- 训练集:3,891;验证集:1,210;测试集:639
-
Weather
- 类别:Overcast, Clear, Foggy, Snowing, Raining
- 训练集:2,798;验证集:1,400;测试集:500
-
Work Zone
- 类别:Work zone, None-work zone
- 训练集:2,121;验证集:1,498;测试集:662
-
Weather-related Road Condition
- 类别:Snowy, Dry, Wet
- 训练集:2,295;验证集:957;测试集:441
-
Time of Day
- 类别:Night, Daytime, Dawn/Dusk
- 训练集:1,656;验证集:1,022;测试集:300
-
Intersection Type
- 类别:None, 4-way, 3-way, Roundabout
- 训练集:1,981;验证集:332;测试集:367
数据集结构
-
目录结构:
dataset/ ├── train/ │ ├── Grade-separated Infrastructure/ │ ├── Road_Function/ │ ├── Weather/ │ ├── Work Zone/ │ ├── Road_Condition/ │ ├── Time of Day/ │ └── Intersection_Type/ ├── val/ └── test/
下载链接
- 下载地址:https://drive.google.com/file/d/1yw4EcfGFGjs2OAa4sWfwQaIDkxDIAR46/view?usp=drive_link
引用
bibtex @misc{li2025acquiringaccumulatingknowledgediverse, title = {Acquiring and Accumulating Knowledge from Diverse Datasets for Multi-label Driving Scene Classification}, author = {Ke Li and Chenyu Zhang and Yuxin Ding and Xianbiao Hu and Ruwen Qin}, year = {2025}, eprint = {2506.17101}, archivePrefix = {arXiv}, primaryClass = {cs.CV}, url = {https://arxiv.org/abs/2506.17101} }




