This project is a fork of gaussian splatting that adds semantic instance information to every single gaussian which is optimized during normal training by doing some extra training.
You can view a demo and mode detailed explaination of how it works here:
