An underwater image dataset for occlusion-aware fish instance segmentation

Underwater visual perception of fish is a core technology in intelligent aquaculture systems. However, occlusion caused by overlapping fish remains a major obstacle to achieving accurate instance segmentation. Robust segmentation under such conditions requires large-scale datasets with both morphological diversity and pixel-level annotations. To this end, we present the Fish Occlusion Dataset (FOD), a large-scale dataset specifically designed for occlusion-aware fish instance segmentation and model evaluation. FOD comprises 14,376 underwater images and 144,894 finely annotated fish instances, categorized into three occlusion levels: whole, part, and fragment, to enable fine-grained performance assessment across varying occlusion scenarios. The dataset includes both original and synthesized images, enabling comprehensive training and evaluation. The data were collected at the Zhuozhou Precision Aquaculture Base of China Agricultural University, and all annotations were manually performed by trained students under expert supervision to ensure consistency and accuracy. We benchmarked eight representative instance segmentation models on the dataset, covering both detection-based and proposal-free architectures. Among them, Mask2Former achieved the highest overall segmentation performance, particularly excelling under severe occlusion.