Fish Occlusion Dataset

The Fish Occlusion Dataset (FOD) is a large-scale image dataset specifically designed for training and evaluating instance segmentation models under challenging underwater occlusion scenarios. It contains 14,376 underwater images and 144,894 pixel-level annotated fish instances. The dataset includes 1,376 real images acquired under controlled conditions, and 13,000 synthetic images generated using compositing techniques to simulate diverse occlusion patterns and fish densities. All fish instances are labeled with both segmentation masks and occlusion-level categories (whole, part, fragment), allowing for fine-grained performance evaluation under different occlusion levels. The dataset follows the MS COCO annotation format and is intended to support the development of robust, segmentation, and occlusion reasoning models in underwater environments.