SISP
收藏资源简介:
SISP数据集是一个专为高分辨率卫星图像中的细粒度船舶实例分割设计的基准数据集。该数据集由10,000张切片图像组成,包含56,693个精细标注的船舶实例,分为四个细粒度类别。所有图像均从SuperView-1卫星收集,分辨率为0.5米。SISP数据集反映了真实卫星场景的特征,如高度类别不平衡、场景多样性、目标密度和尺度的巨大变化,以及高类间相似性和类内多样性,使其更适合实际应用。此外,该数据集还引入了动态特征细化辅助实例分割网络(DFRInst),作为船舶实例分割的基准方法,通过强化关键特征的显式表示来提高船舶实例分割的性能。
The SISP dataset is a benchmark dataset specifically designed for fine-grained ship instance segmentation in high-resolution satellite imagery. This dataset comprises 10,000 cropped satellite images, containing 56,693 finely annotated ship instances categorized into four fine-grained classes. All images were collected from the SuperView-1 satellite, with a spatial resolution of 0.5 meters. The SISP dataset reflects the characteristics of real-world satellite scenes, including severe class imbalance, scene diversity, dramatic variations in target density and scale, high inter-class similarity and intra-class diversity, making it more suitable for practical applications. Additionally, this dataset introduces the Dynamic Feature Refinement Auxiliary Instance Segmentation Network (DFRInst) as the baseline method for ship instance segmentation, which enhances the explicit representation of key features to improve the performance of ship instance segmentation.

- 1SISP: A Benchmark Dataset for Fine-grained Ship Instance Segmentation in Panchromatic Satellite Images中国科学院 · 2024年



