CAW: A realistic benchmark for semi-supervised semantic segmentation in maritime computer vision

Maritime scene understanding is an important task for a variety of industries, yet data collection and annotation for the field are relatively underrepresented and expensive. As a result, current benchmarks for surface-level maritime image segmentation are often smaller scale or comprised of images from different domains. Due to the massive amount of variation in maritime scenes, larger datasets provide greater generalization across multiple domains which is important in safety-critical applications. This paper presents ‘Collection above Water’ (CAW), a novel dataset that aims at addressing these limitations as well as providing a benchmark for semi-supervised image segmentation in the maritime domain. The focus of the dataset is on realistic scenarios using real-world video data to better represent common problems found within maritime computer vision. A systematic review of popular semi-supervised segmentation techniques in the maritime domain is provided. Our results indicate that common failure cases are exacerbated due to the unique properties of the maritime domain having been previously under-explored by other datasets.