Berkeley的大规模自动驾驶视频数据集-BDD100K

发布机构：加州大学伯克利分校 AI 实验室 包含数量：10 万个高清视频序列，10 万张图片 数据格式：标签：.json；图片：.png 图片尺寸：1280*720 数据大小：1.8 TB 数据集中的视频是从美国各地收集的，涵盖不同时间、不同天气条件（包括晴天、阴天和雨天，以及白天和晚上的不同时间）和驾驶场景。 收集数据集的地理位置分布在纽约、伯克利、旧金山等地 数据集中，道路目标检测是为公共汽车、交通灯、交通标志、人、自行车、卡车、摩托车、汽车、火车和乘车人等 100000 张图片上标注 2D 边界框； 实例分割被用于探索具有像素级和丰富实例级注释，相关图像超过 10000 张；引擎区域是从 10 万张图片中学习复杂的可驾驶决策； 车道标记是在 10 万张行车指南图片上的多种车道标注。 车道标记类图片中，标注了实线、虚线、双线、单线等 该数据集由相关论文有《BDD100K: A Diverse Driving Video Database with Scalable Annotation Tooling》，该项目由伯克利 DeepDrive 产业联盟组织和赞助，该联盟研究计算机视觉和机器学习在汽车应用上的最新技术。 资料来源： https://bair.berkeley.edu/blog/2018/05/30/bdd/ 论文地址： https://arxiv.org/pdf/1805.04687.pdf

Publisher: University of California, Berkeley AI Lab
This dataset contains 100,000 high-definition video sequences and 100,000 images.
Data format: Annotations in .json format; Images stored in .png format.
Image resolution: 1280 × 720.
Total data size: 1.8 TB.

The videos in this dataset were collected across the United States, covering diverse time periods, weather conditions (including sunny, cloudy, rainy days, as well as different times of day and night) and driving scenarios. The data collection locations are distributed across New York, Berkeley, San Francisco and other regions.

For road object detection, 2D bounding box annotations are provided on 100,000 images, covering categories such as buses, traffic lights, traffic signs, pedestrians, bicycles, trucks, motorcycles, cars, trains and vehicle passengers.
For instance segmentation, over 10,000 images are annotated with pixel-level and rich instance-level annotations.
For drivable area detection, 100,000 images are utilized to learn complex autonomous driving decision-making.
For lane marking annotation, multiple types of lane annotations are provided on 100,000 driving guidance images, including solid lines, dashed lines, double lines, single lines and other lane categories.

This dataset is associated with the peer-reviewed paper *BDD100K: A Diverse Driving Video Database with Scalable Annotation Tooling*. The project is organized and sponsored by the Berkeley DeepDrive Industrial Alliance, which conducts research on cutting-edge computer vision and machine learning technologies for automotive applications.

Source: https://bair.berkeley.edu/blog/2018/05/30/bdd/
Paper link: https://arxiv.org/pdf/1805.04687.pdf

搜集汇总

背景与挑战

背景概述

BDD100K是由加州大学伯克利分校AI实验室发布的大规模自动驾驶数据集，包含10万个高清视频序列和10万张图片，总大小1.8TB，覆盖美国多地多样化的驾驶场景和天气条件。该数据集提供丰富的标注，如2D边界框、实例分割和车道标记，旨在支持自动驾驶计算机视觉和机器学习研究。