Computer vision techniques such as deep learning have quickly become of interest to ecologists for automatically processing large volumes of video and image-based data. However, training deep learning models often require large volumes of manually annotated footage to create a robust and accurate model. The collection and annotation of these training datasets can incur high initial labour cost and may not be feasible for some research projects. The accessibility of publicly available datasets that are pre-annotated for easy implementation is imperative for continued research and understanding of computer vision technology as a viable method to automate the processing of visual data. In this report, we provide a dataset containing ~ 9,000 annotated fish in unconstrained conditions in a key coastal habitat - seagrass meadows - collected via remote underwater video. These images include object instance annotations which consist of a corresponding image, label, bounding box and segmentation mask. These data can be used for training several different computer vision models and for investigating the effects of pre- or post-processing steps to improve model performance when predicting data in awuatic habitats. The purpose of this report, in conjunction with the annotated dataset, is to advance the use of CV techniques and further the growth in labelled fish datasets publicly available.