dense500
收藏资源简介:
dense500是一个专为多尺度卫星图像合成任务设计的基准数据集,由华盛顿大学圣路易斯分校的研究团队构建。该数据集包含500个完全观测的深度4四叉树,每个四叉树由85个连续缩放级别的256×256像素瓦片组成,覆盖五个缩放窗口,分别对应100个地理上独特的站点。创建过程严格遵循人口密度加权、大陆分层和至少80公里空间间隔的采样策略,并剔除了与训练数据重叠及影像源不一致的站点,以确保评估的独立性。dense500旨在为多尺度瓦片补全任务提供标准化评估基准,推动生成模型在垂直和水平维度上的一致性与无缝性,从而解决稀疏种子瓦片生成全局一致多分辨率金字塔的挑战。
Dense500 is a benchmark dataset specifically designed for the multi-scale satellite image synthesis task, constructed by the research team from Washington University in St. Louis. This dataset contains 500 fully observed depth-4 quadtrees, each consisting of 256×256 pixel tiles across 85 consecutive scaling levels, covering five scaling windows that correspond to 100 geographically unique sites respectively. Its construction strictly follows sampling strategies including population density weighting, continental stratification, and a spatial separation of at least 80 kilometers, while excluding sites that overlap with training data or have inconsistent image sources to ensure the independence of evaluation. Dense500 aims to provide a standardized evaluation benchmark for the multi-scale tile inpainting task, promote the consistency and seamlessness of generative models across vertical and horizontal dimensions, thereby addressing the challenge of generating globally consistent multi-resolution pyramids from sparse seed tiles.




