The authors present a novel approach to 3D Gaussian Splatting by viewing the set of Gaussians as a random sample from an underlying probability distribution, akin to Markov Chain Monte Carlo samples. By introducing noise and utilizing Stochastic Gradient Langevin Dynamics, they eliminate the need for carefully engineered cloning and splitting strategies. The method provides higher quality renderings, allows for easy control over the number of Gaussians, and is robust to initialization challenges. Results on standard evaluation scenes demonstrate the effectiveness of the approach in improving rendering quality and preserving scene details. The proposed method offers a significant departure from conventional Gaussian splatting techniques.
https://ubc-vision.github.io/3dgs-mcmc/