GaussUnveil: Unified Occlusion-Aware Gaussian Refinement for Sparse-View Scene Reconstruction

Guoqing Wang, Pin Tang, Xiangxuan Ren, Liping Hou, Chao Ma

Keywords: Occlusion-aware, Sparse-View Scene Reconstruction, Gaussian Refinement

TL;DR: We propose GaussUnveil, an occlusion-aware selective-refinement framework that shifts the paradigm from refining everywhere to refining where it matters.

Abstract:

Ego-centric 3D reconstruction from sparse, low-overlap views is challenging, as cross-view correspondences are limited, occlusions occur frequently, and per-camera frusta often truncate scene structures. Explicit Gaussian pipelines mitigate some of these challenges, and the dual-branch methods that couple pixel- and volume-based Gaussians (e.g., Omni-Scene) further enhance robustness. However, they typically refine large numbers of Gaussians uniformly, regardless of visibility or structural ambiguity.
We propose GaussUnveil, an occlusion-aware selective-refinement framework that shifts the paradigm from refining everywhere to refining where it matters. By unveiling regions of uncertainty near occlusions, GaussUnveil identifies where additional Gaussian refinement is needed. Specifically, we derive occlusion masks from depth-gradient discontinuities, lift them into the 3D volume to initialize a compact set of Gaussian queries. Then, we employ a lightweight refinement block that aggregates self-context and multi-view features while iteratively updating the mean and covariance of each Gaussian query under differentiable rendering. Extensive experiments on both ego-centric and scene-centric benchmarks demonstrate the effectiveness of the proposed method compared to the state-of-the-art reconstruction methods. For instance, GaussUnveil delivers superior performance while using about 30\% fewer Gaussians and is approximately 34\% faster than Omni-Scene.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 3369