SHIFT

我们介绍了最大的用于自动驾驶的多任务综合数据集SHIFT。它呈现云量，雨雾强度，一天中的时间以及车辆和行人密度的离散和连续变化。SHIFT具有针对多个主流感知任务的全面传感器套件和注释，可以在不断提高的域移位水平下研究感知系统性能的下降，从而促进了持续适应策略的发展，以缓解此问题并评估模型的鲁棒性和通用性。我们的数据集和基准工具包可在www.vis.xyz/shift上公开获得。

We introduce SHIFT, the largest multi-task comprehensive dataset for autonomous driving. It encompasses discrete and continuous variations in cloud cover, rain/fog intensity, time of day, as well as vehicle and pedestrian densities. SHIFT provides comprehensive sensor suites and annotations for multiple mainstream perception tasks, allowing researchers to investigate the performance degradation of perception systems under progressively increasing levels of domain shift. This facilitates the development of continuous adaptation strategies to mitigate such issues, as well as the evaluation of model robustness and generalizability. Our dataset and benchmark toolkit are publicly accessible at www.vis.xyz/shift.

搜集汇总

背景与挑战

背景概述

SHIFT是一个专为自动驾驶设计的大规模多任务综合数据集，通过模拟云量、雨雾强度、时间变化及交通密度等多种域移位场景，支持多个感知任务和传感器套件，旨在研究模型在域变化下的性能下降和鲁棒性，促进持续适应策略的发展。该数据集由Google、苏黎世联邦理工学院和慕尼黑工业大学于2022年发布，采用CC BY-SA 4.0许可证开放访问。