SAT-MTB-SOS: A Benchmark Dataset for Satellite Video Single Object Segmentation

Satellite video object segmentation is capable on various applications such as specific events monitoring and outline moving details by continuously observing large areas. However, the small size of targets and appearance deformation make the Satellite video object segmentation more difficult, and the lack of satellite video datasets has limited the development of satellite video object segmentation. Therefore, it is necessary to establish a benchmark to fill the vacancies and advance the research. In this work, we present SAT-MTB-SOS, the first frame-by-frame and pixel-level annotated benchmark dataset for satellite video single object segmentation. SAT-MTB-SOS contains 113 high-quality video sequences with a total of 13,500 frames, covering five categories of typical targets in satellite videos: airplanes, trains, vehicles, ships, and buildings. SAT-MTB-SOS includes various video object segmentation challenges such as occlusion, motion blur, and non-rigid variations. Moreover, we conduct detailed evaluation and analysis of 15 representative video object segmentation algorithms, which provides a basis for further research on satellite video object segmentation