Hi @shan,
I worked on the mentioned GSoC project. With the latest update of the model generator, you can attach a cam to the boat, take some images (the images should be clean, like the demo video in the post: no high splash, no significant boat parts) and generate your own segmentation model following the instruction of a GUI. It’s not out of the box but it might be the best solution due to the fact that we are lack of both datasets and AI camera’s computing capability. If one day we can put a RTX40xx level card onto the boat, there won’t be need to distillate meta’s SegmentAnythingModel anymore. 
Also, the obstacle detection part is a bit rough (simply based on detection grids), but when you set the grid num to the same as pixel num, it actually outputs obstacle-only point cloud to ArduPilot. Then traditional SLAM, etc. can be introduced maybe.
And indeed, you may need some tweaks like adjusting the confidence thresthold before it works. It’s sad, but it’s true. 
Chenghao