mini-weather dataset
收藏资源简介:
WeatherGen是一个用于生成多样天气LiDAR数据的统一框架,由南京理工大学PCA实验室提出。该框架包括一个基于地图的数据生成器MDP,能够提供用于模型预训练的多样天气LiDAR数据。WeatherGen通过设计一种新的生成器——蜘蛛曼巴生成器SMG,有效保持了LiDAR数据的物理结构。此外,通过对比学习机制,设计了一个控制器CLC,用于生成具有紧凑语义知识的控制信号。利用WeatherGen构建的mini-weather数据集,应用于三维目标检测任务,证明了生成数据可以促进在恶劣天气条件下的下游任务性能。
WeatherGen is a unified framework for generating diverse weather LiDAR data, proposed by the PCA Laboratory of Nanjing University of Science and Technology. This framework incorporates a map-based data generator (MDP) that can provide diverse weather LiDAR data for model pre-training. WeatherGen designs a novel generator, the Spider Mamba Generator (SMG), which effectively preserves the physical structure of LiDAR data. Additionally, leveraging contrastive learning mechanisms, a Contrastive Learning Controller (CLC) is designed to generate control signals with compact semantic knowledge. The mini-weather dataset constructed using WeatherGen is applied to 3D object detection tasks, and the results verify that the generated data can enhance the performance of downstream tasks under adverse weather conditions.




