This is a research prototype and benchmark effort, not a claim of inventing multiview panorama repair. The following approaches motivate future work; none is silently substituted into the reported baseline.
COLMAP is the classical structure-from-motion and multiview-stereo route: estimate cameras and scene points, then denser surfaces. It provides the right geometric abstraction but depends on texture, overlap and appropriate camera models. PAVER currently fits a shared road plane, which cannot explain a full street scene.
VGGT and MapAnything predict multiview geometry and cameras. PanoVGGT, with released code, is especially relevant: it processes panoramic images for poses, depth and point clouds. Its PanoCity dataset is synthetic; training overlap must be excluded from any independent evaluation. CUDA-oriented dependencies do not imply it runs on this project's MPS hardware. No checkpoint has been evaluated here.
PaGeR estimates panoramic depth and normals. ZüriPano provides real outdoor LiDAR-based geometry references; it is not automatically a contiguous donor network benchmark. MASt3R-SLAM combines learned point maps with tracking, fusion and loop closure. Argus studies sparse panoramic reconstruction and covisibility in indoor settings; road transfer remains to be established.
The useful experiment is to infer a shared surface map from allowed observations, project a missing target ray onto it, test which source camera can actually see that surface, and transfer only adequately sampled observed pixels. Depth/pose uncertainty, occlusion checks and appearance changes must accompany coverage.
Multiview inpainting work from 2016 and Street View privacy protection, CVPR 2019 are direct prior art for urban multiview repair. Street Gaussians separates static streets and dynamic objects for novel-view rendering. GS-RoadPatching uses similar Gaussian patches to repair road appearance; a plausible substitute is not necessarily the previously hidden physical surface. RoadGaussian uses masks and image inpainting in a Gaussian reconstruction pipeline. SPIn-NeRF and Geometry-Aware Diffusion for Multiview Scene Inpainting address consistent scene completion. These can be useful when no donor saw the surface, but synthesized content must remain distinguishable from evidence.
For PAVER, keep two outcomes separate: observed-surface recovery, with source provenance and calibrated acceptance tests, and plausible completion, which cannot be scored as recovery without independent truth. Current experiments implement the first, restricted to a road plane, and retain abstentions.