Coral reefs are reported to support up to a quarter of all marine life on Earth. Coral reefs are also reported to be undergoing a severe decline in the last decades worldwide. It is therefore very important to continuously monitor the health of coral reefs. However, this is a challenging task. On the one hand the quality of underwater images can be affected by the physical and chemical characteristics of underwater conditions. On the other hand, it is unfeasible for human scientists to manually review and analyze hundreds or even thousands of underwater images and videos. Recently, Deep Learning methods have been used to facilitate this task. Deep Learning methods, however, pose challenges of their own, namely the need to derive good quality training datasets from new unlabeled data. This work proposes an end-to-end pipeline that receives as input new unlabeled data and outputs a fully trained model capable of performing coral and fish segmentation in underwater videos, with as little human intervention as possible. An important characteristic of this pipeline is its ability to derive a good quality training dataset in an easier, faster and interactive way via assisted labeling with a foundation model. The use of image restoration techniques to improve the quality of the images and computer vision techniques to refine the annotations are also contributing factors to the good quality of the derived training dataset. The promising results achieved seem to validate the quality of the pipeline proposed.