Unsupervised Llamas

通过创建用于自动驾驶的高清地图 (包括基于激光雷达的车道标记) 来注释无监督的Llamas数据集。可以根据这些地图对自动车辆进行定位，并将车道标记投影到相机框架中。通过最小化图像中已经检测到的标记与投影标记之间的差异来优化3D投影。通过使用更好的检测器，优化差异度量并添加一些时间一致性，可以实现进一步的改进。超过100,000个带注释的图像，超过100米个的注释，1276x717个像素的分辨率。

We annotate the unsupervised Llamas dataset by generating high-definition (HD) maps for autonomous driving, which include LiDAR-based lane markings. These maps can be used to localize autonomous vehicles and project lane markings onto the camera frame. 3D projection is optimized by minimizing the discrepancy between the lane markings detected in images and the projected ones. Further improvements can be achieved by adopting more robust detectors, optimizing the discrepancy metric, and incorporating temporal consistency. The dataset contains over 100,000 annotated images, over 1,000,000 annotations, and a resolution of 1276 × 717 pixels.

搜集汇总

背景与挑战

背景概述

该数据集由Bosch Automated Driving于2019年发布，专注于自动驾驶领域，通过创建高清地图来注释车道标记，包含超过10万张标注图像和1亿个注释，图像分辨率为1276x717像素。