Rendering-Aware Bayesian 3D Gaussian Splatting with Native Uncertainty and Adaptive Complexity Control

· Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision & Graphics · Depth: Expert, extended

Summary

A new rendering-aware Bayesian framework enhances 3D Gaussian Splatting (3DGS) by integrating native uncertainty and adaptive complexity control. This approach addresses standard 3DGS limitations, particularly under limited view support, by tracking geometry with a Normal-Inverse-Wishart (NIW) posterior and optionally using a Dirichlet-process extension for probabilistic component usage. In a fixed-budget 16-to-32 active-view selection task, NIW acquisition improves PSNR by +0.453 dB and LPIPS by -0.0146 over a standard-ensemble baseline, winning 29/39 scene-seed pairs. NIW native intervals reduce 95% coverage error by approximately 17x compared to a shared proxy and are about 10x closer to nominal coverage than a 3-member deep ensemble at one-third the training cost. The method achieves a +0.030 dB PSNR gain over standard 3DGS with only 1.6% additional training time, establishing it as a practical probabilistic scene representation.

Key takeaway

For AI Scientists and Machine Learning Engineers developing 3D Gaussian Splatting applications, especially those involving active view selection or operating with limited data, you should consider integrating this rendering-aware Bayesian 3DGS. Its native uncertainty quantification significantly improves view acquisition decisions, yielding higher reconstruction quality and providing empirically calibrated prediction intervals. This approach offers a principled way to enhance model robustness and transparency for decision-facing tasks, with only a 1.6% training overhead compared to standard 3DGS.

Key insights

Bayesian 3DGS integrates native uncertainty and adaptive complexity for robust novel-view synthesis and active view selection.

Principles

Method

Tracks 3DGS geometry with a Normal-Inverse-Wishart posterior using renderer-derived surrogate summaries. Re-renders posterior samples to generate native per-pixel predictive uncertainty.

In practice

Topics

Best for: Research Scientist, AI Scientist, Computer Vision Engineer, Machine Learning Engineer

Related on AIssential

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.