Cross-spectral Gaussian Splatting with implicit surface consistency

Cross-spectral scene reconstruction from asynchronous or imperfectly synchronized multi-sensor observations remains challenging because spectral appearance differences complicate cross-camera correspondence, residual timing offsets can introduce view-dependent pose discrepancies during camera motion, and photometric optimization alone does not guarantee a common scene surface. Building on our previous cross-spectral Gaussian framework for per-view pose recovery and shared Gaussian geometry, this work focuses on the remaining geometric inconsistency of the unified representation. The implemented framework couples an explicit Gaussian Splatting representation with a spectrum-independent neural signed distance field and is applied to cross-spectral scene reconstruction, camera pose estimation, and novel-view synthesis. Depth and normal cues rendered from the Gaussian representation supervise the shared implicit surface, while its zero-level set anchors Gaussian centers and guides subsequent densification. Monocular depth priors further stabilize sparsely initialized regions. The implicit branch is used only during optimization, preserving explicit real-time Gaussian rasterization at inference. Experiments demonstrate that the proposed extension primarily improves non-reference spectral alignment and surface reconstruction while maintaining comparable reference-spectrum rendering quality. The denser geometrically refined representation introduces only a small reduction in rendering speed. Comparisons with recent pose-free and Gaussian-based surface reconstruction methods, together with component-wise ablations using cross-view depth consistency, further validate the complementary roles of pose refinement, surface anchoring, monocular depth supervision, and guided densification.