DP-Splat: Bayesian Nonparametric Complexity Control for Gaussian Splatting

Research Approach

This paper introduces DP-Splat, a framework that reframes 3D Gaussian Splatting as a Bayesian nonparametric inference problem. Traditional Gaussian splatting relies on heuristic "adaptive density control" (splitting and pruning) to manage the number of components. DP-Splat replaces these heuristics with a truncated stick-breaking Dirichlet-process (DP) prior or a sparse overfitted finite Dirichlet prior. By casting the fitting process as conjugate variational inference, the model allows the number of occupied components to emerge naturally from the data. The authors provide a closed-form coordinate-ascent variational inference (CAVI) algorithm and a stochastic natural-gradient variant that scales to millions of points on standard hardware.

Key Findings

• Complexity Selection: The DP prior acts as an effective complexity controller. In empirical tests, DP-Splat recovers the true number of components on synthetic data and achieves competitive rendering quality on 3D scenes with significantly fewer components (5.9–7.6x fewer) than fixed-budget baselines.
• Statistical Rigor: The authors provide a corrected, rigorous truncation-error bound for the DP prior, demonstrating that commonly used exponential approximations are anti-conservative in the small-alpha regime where splatting models typically operate.
• Variational Dynamics: The study documents a significant gap between theoretical posterior asymptotics and variational practice. While theory suggests DP priors should inflate component counts, the coordinate-ascent dynamics in practice often lead to component death, causing the DP and sparse finite mixture behaviors to invert compared to their expected theoretical trajectories.
• Calibration: The model provides well-calibrated posterior-predictive color uncertainty as a byproduct of the Bayesian formulation, a feature missing in standard fixed-precision splatting approaches.