ONCE
收藏资源简介:
ONCE数据集是由复旦大学信息科学与技术学院和上海人工智能实验室共同创建的大规模点云数据集,专为自动驾驶领域的预训练任务设计。该数据集包含约100万条数据,涵盖多种场景和天气条件,旨在通过丰富的数据分布学习可泛化的表示。数据集的创建过程中,采用了类别感知的伪标签生成策略和多样性预训练处理器,以增强数据的场景和实例级多样性。此数据集的应用领域主要集中在自动驾驶相关的感知任务,旨在解决自动驾驶系统在不同环境和条件下的泛化能力问题。
ONCE Dataset is a large-scale point cloud dataset co-developed by the School of Information Science and Technology, Fudan University and Shanghai AI Laboratory, specifically designed for pretraining tasks in the autonomous driving domain. It contains approximately 1 million data samples, covering various scenarios and weather conditions, aiming to learn generalizable representations through rich data distributions. During the dataset creation process, a category-aware pseudo-label generation strategy and a diverse pretraining processor were adopted to enhance the data diversity at both scenario and instance levels. The main application fields of this dataset focus on perception-related tasks in autonomous driving, with the goal of solving the generalization capability issues of autonomous driving systems under different environments and conditions.




