Voxel51/SkyScenes

SkyScenes是一个用于航空场景理解的综合合成数据集，已接受为ECCV 2024的论文。该数据集包含33,600张使用CARLA模拟器从无人机视角捕获的航空图像。每个图像都有密集的像素级注释，包括语义分割、实例分割和深度信息。数据集包含了不同的城镇布局、天气/时间条件、视角组合以及不同高度和俯仰角度的变化。此存储库中包含的数据集是完整数据集的一个子集，仅包含特定的天气条件、城镇布局、高度和俯仰角度。数据集适用于训练和评估航空场景理解模型，支持多模态分割模型的发展，并可用于研究合成到现实域的自适应以及模型在不同视角条件下的行为。

SkyScenes is a comprehensive synthetic dataset for aerial scene understanding that was recently accepted to ECCV 2024. The dataset contains 33,600 aerial images captured from UAV perspectives using the CARLA simulator. Each image comes with dense pixel-level annotations for semantic segmentation, instance segmentation, and depth information. The dataset includes variations in town layouts, weather/time conditions, viewpoint combinations, as well as different heights and pitch angles. The dataset subset in this repository contains specific weather conditions, town layouts, heights, and pitch angles. It is suitable for training and evaluating aerial scene understanding models, supports the development of multi-modal segmentation models, and can be used for research on synthetic-to-real domain adaptation and model behavior under different viewing conditions.