Analyzing and understanding marine imagery has gained increasing attention within both computer vision and marine communities. It is challenging to build a foundation model for marine visual analysis. The scarcity of labeled data is the most hindering issue, and marine photographs illustrate significantly different appearances and contents from general in-air images. Based on our previous work MarineGPT, the first vision-language model specifically on the marine domain with extensive marine knowledge, our team created MarineInst20M, the largest marine image dataset to date, with 2.42 million images and 19.2 million masks in total and introduced MarineInst, a foundation model for marine visual analysis which can segment and describe the marine object instances. The dataset and model support a wide range of marine visual analysis tasks, from image-level scene understanding to regional mask-level instance understanding. Besides, the model exhibits strong generalization ability and flexibility to support various downstream tasks with state-of-the-art performance.