LOGAussian: Efficient Local Gathering for Online Feed-forward 3DGS

Meng Wei, Cheng Zhang, Qianyi Wu, Jianmin Zheng, Hamid Rezatofighi, Jianfei Cai

Keywords: neural rendering; 3DGS; feed-forward 3dgs; 3d reconstruction;

Abstract:

Feed-forward 3D Gaussian Splatting (3DGS) has attracted increasing attention for its broad applicability and real-time inference capabilities. Despite recent progress through advanced backbones, prevailing pipelines remain offline: concatenating per-view, pixel-aligned splats introduces redundancy and, under streaming input, accumulates errors over time. To address this, we propose \textbf{LOGAussian} (LOcal GAthering Gaussians), a lightweight post-hoc module that maintains an incrementally updated and render-ready global 3D Gaussian representation from sequential posed images without per-scene optimization. The approach integrates global scene context and models local correlations among predicted Gaussians, explicitly accommodating sparse-view inputs with depth noise and geometric imprecision.
With about 1\% additional parameters, our approach yields a compact, consistent set of splats while maintaining or improving rendering quality as the stream progresses.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 5871