nuScenes and VoD

The dataset comprises large-scale, high-resolution multimodal sensor data for 3D object detection, including LiDAR point clouds, synchronized RGB images, and depth maps. It contains ​XX,000​ annotated scenes captured across diverse environments (urban, highway, indoor) under varying lighting and weather conditions. Each scene is densely annotated with 3D bounding boxes for ​XX categories​ (e.g., vehicles, pedestrians, cyclists), with precise localization (x, y, z), orientation (yaw, pitch, roll), and occlusion/truncation labels. The dataset is split into training (XX%), validation (XX%), and test (XX%) sets, with the test set labels withheld to ensure unbiased benchmarking. Challenges include long-tail class distributions, dynamic occlusions, and sparse point cloud regions, reflecting real-world complexity. Auxiliary metadata (calibration parameters, timestamps) and optional scene-level semantic segmentation labels are provided to support multi-task learning.

本数据集面向三维目标检测任务，提供大规模高分辨率多模态传感器数据，涵盖激光雷达（LiDAR）点云、同步RGB图像以及深度图。其包含XX,000个标注场景，采集自城市、高速路、室内等多样化环境，且覆盖不同光照与天气条件。每个场景均针对XX个类别（例如车辆、行人、骑行者）进行稠密三维标注，包含精准的位置信息（x、y、z坐标）、姿态参数（偏航、俯仰、滚转角）以及遮挡/截断标签。该数据集划分为训练集（XX%）、验证集（XX%）与测试集（XX%），其中测试集标签未予公开，以确保基准测试的公正性。该数据集存在长尾类别分布、动态遮挡以及点云稀疏区域等挑战，贴合真实应用场景的复杂性。此外还提供校准参数、时间戳等辅助元数据，以及可选的场景级语义分割标签，可用于支持多任务学习。