CurveLanes

CurveLanes是一个新的基准车道检测数据集，具有150K个车道图像，适用于交通车道检测中的曲线和多车道等困难场景。它是在中国多个城市的真实城市和高速公路场景中收集的。它是迄今为止最大的车道检测数据集，并为社区建立了更具挑战性的基准。
我们将整个数据集150K分为三个部分: train:100K，val: 20K和testing: 30K。该数据集中的大多数图像的分辨率为2650 × 1440。
对于每个图像，我们使用自然的三次样条线手动注释图像中的所有通道。所有图像都经过精心选择，以使大多数图像至少包含一个曲线车道。在这个数据集中可以找到更困难的场景，例如s形曲线、Y形车道、夜间和多车道 (车道线的数量超过4条)。

CurveLanes is a novel benchmark lane detection dataset consisting of 150K lane images, tailored for challenging traffic lane detection scenarios including curved lanes and multi-lane configurations. It is collected from real urban and highway scenes across multiple cities in China, and stands as the largest lane detection dataset to date, establishing a more demanding benchmark for the global research community.
We split the entire 150K dataset into three subsets: 100K samples for training, 20K samples for validation, and 30K samples for testing. Most images in this dataset have a resolution of 2650 × 1440.
For each image, all lane markings are manually annotated using natural cubic splines. All images are carefully curated such that the majority contain at least one curved lane. A wider range of challenging scenarios are included in this dataset, such as S-shaped curves, Y-shaped lanes, nighttime driving scenes, and multi-lane situations where the number of lane lines exceeds four.

搜集汇总

背景与挑战

背景概述

CurveLanes是一个专注于车道检测的大规模数据集，包含150K张高分辨率图像，特别针对曲线和多车道等复杂交通场景进行标注。该数据集由华为诺亚方舟实验室和中山大学于2020年发布，旨在为自动驾驶领域提供更具挑战性的基准测试资源。