iSAID

现有的地球视觉数据集适用于语义分割或对象检测。iSAID是第一个用于在航空图像中进行实例分割的基准数据集。这个大规模且带有密集注释的数据集包含2,806个高分辨率图像中15个类别的655,451个对象实例。iSAID的独特特征如下 :( a) 具有高空间分辨率的大量图像，(b) 15个重要且常见的类别，(c) 每个类别的大量实例，(d) 每个图像的大量标记实例，这可能有助于学习上下文信息，(e) 巨大的物体尺度变化，通常在同一图像内包含小，中和大物体，(f) 图像内具有不同方向的物体的不平衡和不均匀分布，描绘了现实生活中的空中条件，(g) 几个小尺寸的物体，具有模棱两可的外观，只能通过上下文推理来解决，(h) 由专业注释器执行的精确实例级注释，由遵循明确定义的指南的专家注释器进行交叉检查和验证。

Existing Earth vision datasets are primarily developed for semantic segmentation or object detection. iSAID is the first benchmark dataset for instance segmentation in aerial imagery. This large-scale, densely annotated dataset contains 655,451 object instances belonging to 15 categories across 2,806 high-resolution images. The unique characteristics of iSAID are as follows: (a) A large volume of images with high spatial resolution; (b) 15 important and prevalent object categories; (c) A substantial number of instances per category; (d) A large quantity of labeled instances per image, which can aid in learning contextual information; (e) Extreme variations in object scale, with small, medium and large objects frequently coexisting within a single image; (f) An imbalanced and non-uniform distribution of objects with varying orientations within the images, which accurately reflects real-world aerial scenarios; (g) Several small-sized objects with ambiguous appearances that can only be correctly identified via contextual reasoning; (h) Precise instance-level annotations performed by professional annotators, which are subsequently cross-checked and validated by expert annotators adhering to well-defined guidelines.

搜集汇总

背景与挑战

背景概述

iSAID是一个专为航空图像实例分割设计的大规模数据集，包含2,806张高分辨率图像和15个类别的655,451个对象实例，具有高空间分辨率、物体尺度变化大和方向分布不均等特点。