Abstract
Fish farming is a promising economic activity that promotes the social development, protects the ecological environment, and enhances the quality of human life. In recent years, various computer vision models have been established for assessing aquaculture density and monitoring fish health. However, existing datasets are generally characterised by larger fish sizes and low density, making them unsuitable for detecting small targets such as fish fry. This paper presents a dataset comprising 1101 images of largemouth bass (Micropterus salmoides) fry, specifically designed for small target detection in dense scenes. Each image contains a variable number of fish fries, ranging from 20 to 80 individuals. To facilitate health assessment in the aquaculture, a small number of dead fish fries are included in each image. The entire dataset is annotated with a total of 51,119 live fish fry and 3586 dead ones. Additionally, among the 80 images depicting high-density scenarios, there are complex situations such as overlap, occlusion, and adhesion, which pose challenges to the small target detection task. The dataset is annotated using the Labelimg tool and converted to the COCO format. It can be applied to a variety of scenarios, including seedling rearing, fry retailing, and survival assessments. It is also valuable for biomass estimation and aquaculture density control applications. In summary, this dataset provides an invaluable resource for the research community, advancing studies on fry counting and fish population health, thus contributing to the development of intelligent aquaculture.