基于互联网规模生态数据的视觉搜索数据集
收藏资源简介:
该数据集是基于互联网规模生态数据构建的视觉搜索数据集,旨在解决在卫星图像中目标不可见的情况下进行视觉搜索的问题。数据集包含437k张训练图像和4k张验证图像,图像均标记有多个未见分类目标的坐标。该数据集的创建过程使用了Sentinel-2卫星图像和iNat2021数据集,并利用了生物分类法的层次结构,以促进在不同层次上进行基准评估。数据集主要用于评估Search-TTA框架的性能,该框架能够在搜索过程中动态地改进视觉模型的预测。
This dataset is a visual search dataset constructed using internet-scale ecological data, aimed at addressing the visual search problem where targets are not visible in satellite imagery. It contains 437k training images and 4k validation images, with each image annotated with coordinates of multiple targets from unseen categories. Developed with Sentinel-2 satellite imagery and the iNat2021 dataset, this dataset utilizes the hierarchical structure of biological taxonomy to enable benchmark evaluation across various taxonomic levels. It is primarily employed to evaluate the performance of the Search-TTA framework, which can dynamically improve the prediction outputs of visual models during the search procedure.



