Generalized gaussian splatting (independent research). Winter 2023
Though gaussian splatting has achieved great results, the smooth nature of gaussians makes it tend towards having fuzzy edges, or requiring many gaussians at object boundaries. This work presents a solution: instead of using gaussians, use generalized gaussians. This allows for a learned ‘shape’ parameter, controlling the sharpness of the fall-off. With the same number of gaussians and the same number of training steps, generalized gaussian splatting achieved far sharper results.
