PanoHK360
收藏资源简介:
PanoHK360是一个用于深度估计的大规模、高分辨率城市全景数据集与基准。该数据集采集自香港密集的城市环境,包含连续行驶序列捕获的户外街景,涵盖街道峡谷、交叉路口、建筑立面等复杂场景。其核心特征在于通过传感器直接测量提供高精度度量深度信息。数据通过集成Teledyne FLIR Ladybug全景相机和RIEGL VUXR-1HA22测绘级激光雷达扫描仪的车载多传感器平台采集,并进行了时间同步,确保RGB图像与几何数据精确对应。数据以等距柱状投影(ERP)格式提供,分辨率高达8000×4000像素(8K)。每个数据帧包含多模态标注:RGB全景图、源自激光雷达投影的度量深度图、表面法线图、原始点云数据以及6自由度相机位姿。深度标注基于真实激光雷达回波数据生成,保证了几何真实性,而非合成渲染或神经网络预测的伪标签。发布版本为“R101 20230413--filter”,以压缩包和可浏览文件夹形式提供,包含逐帧的RGB、深度、法线、位姿和点云文件。该数据集旨在支持全景深度估计模型的即插即用式训练和评估,特别适用于ERP感知架构和360度几何感知方法的研究,同时也适用于3D重建、自动驾驶等计算机视觉任务。数据集采用CC BY 4.0许可协议发布。
PanoHK360 is a large-scale, high-resolution urban panoramic dataset and benchmark for depth estimation. It is collected from the dense urban environment of Hong Kong, containing outdoor street scenes captured in continuous driving sequences, covering complex scenarios such as street canyons, intersections, and building facades. Its core feature lies in providing high-precision metric depth information directly measured by sensors. Data is acquired through a vehicle-mounted multi-sensor platform integrating a Teledyne FLIR Ladybug panoramic camera and a RIEGL VUXR-1HA22 surveying-grade LiDAR scanner, with time synchronization ensuring precise correspondence between RGB images and geometric data. The data is provided in equirectangular projection (ERP) format with a resolution of up to 8000×4000 pixels (8K). Each data frame includes multimodal annotations: RGB panoramic images, metric depth maps derived from LiDAR projection, surface normal maps, raw point cloud data, and 6-degree-of-freedom camera poses. Depth annotations are generated based on real LiDAR echo data, ensuring geometric authenticity rather than synthetic rendering or pseudo-labels from neural network predictions. The release version is R101 20230413--filter, provided in both compressed packages and browsable folders, containing frame-by-frame RGB, depth, normal, pose, and point cloud files. This dataset aims to support plug-and-play training and evaluation for panoramic depth estimation models, particularly suitable for research on ERP perception architectures and 360-degree geometric perception methods, as well as computer vision tasks such as 3D reconstruction and autonomous driving. The dataset is released under the CC BY 4.0 license.




