Skip to content
Book Open access

Uncertainty-Aware Consensus 3D Gaussian Splatting for High-Fidelity Lensless Novel View Synthesis

Nov 2026 · Proceedings of the 34th ACM International Conference on Multimedia · 0 citations · 67 references

Abstract

Lensless imaging has emerged as a promising paradigm for ultra-compact, low-cost, and privacy-preserving visual sensing. However, extending recent advances in 2D lensless reconstruction to 3D novel view synthesis remains challenging. Existing methods mainly rely on neural-radiance-field-based representations, which suffer from slow rendering and limited practicality. In this work, we take the first step toward high-fidelity lensless novel view synthesis using 3D Gaussian Splatting (3DGS). This setting poses three major challenges: maintaining cross-view geometric consistency, suppressing ghosting in occluded regions, and restoring high-frequency textures under efficient optimization. To address these issues, we propose Uncertainty-Aware Consensus 3D Gaussian Splatting (UACGS), a framework that combines structure-aware consensus preprocessing with mask-guided optimization. Specifically, our framework incorporates: (1) a structure-aware consensus module that leverages adjacent camera poses to filter out occlusion-induced erroneous projections; (2) an uncertainty-aware mask-guided supervision mechanism that effectively corrects reconstruction errors caused by physical limitations; and (3) a contextual loss design that restores fine-grained high-frequency textures while preserving geometric stability. Extensive experiments show that our framework achieves superior reconstruction quality across datasets, while also delivering strong cross-dataset generalization, real-time rendering, and highly efficient training. The code is publicly available at https://github.com/YUNYU7472/UACGS.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.