MAS3K: An Open Dataset for Marine Animal Segmentation

Recent advances in marine animal research have raised significant demands for fine-grained marine animal segmentation techniques. Deep learning has shown remarkable success in a variety of object segmentation tasks. However, deep based marine animal segmentation is lack of investigation due to the short of a marine animal dataset. To this end, we elaborately construct the first open Marine Animal Segmentation dataset, called MAS3K, which consists of more than three thousand images of diverse marine animals, with common and camouflaged appearances, in different underwater conditions, such as low illumination, turbid water quality, photographic distortion, etc. Each image from the MAS3K dataset has rich annotations, including an object-level annotation, a category name, an animal camouflage method (if applicable), and attribute annotations. In addition, based on MAS3K, we systematically evaluate 6 cutting-edge object segmentation models using five widely-used metrics. We perform comprehensive analysis and report detailed qualitative and quantitative benchmark results in the paper. Our work provides valuable insights into the marine animal segmentation, which will boost the development in this direction effectively.