Multi-view context-consistent image reconstruction for building facades

Accurate building facade understanding supports urban analytics, hazard modeling, and robotic infrastructure inspection, but suburban facades are frequently obstructed by vegetation, parked vehicles, and fences. Street-level imagery (SLI), including Google Street View (GSV), offers broad coverage, yet existing single-view inpainting cannot exploit observations of the same facade from other viewpoints and may generate plausible but unsupported structure. This raises the research question addressed in this work: can multi-view facade reconstruction be extended from sensor-rich mobile mapping to RGB-only street-view imagery while preserving geometrically consistent facade evidence and controlling unsupported generative completion? We present a sensor-light extension of a previously introduced sensor-rich multi-view reconstruction pipeline. The RGB pathway removes the LiDAR requirement by combining perspective views rendered from street-view panoramas, explicitly derived pinhole intrinsics, monocular metric depth (UniDepthV2), cross-view depth-scale correction, conditional local pose refinement, and evidence-gated multi-view fusion before diffusion inpainting. Quantitative evaluation on LiDAR-paired mobile sequences with LiDAR withheld from the sensor-light input shows that scale correction reduces depth AbsRel from 0.142 to 0.091 and median reprojection error from 3.8 to 2.4 pixels, while increasing downstream door/stoop macro-F1 from 0.82 to 0.86, compared with 0.88 for the LiDAR reference. The scale-corrected sensor-light configuration reaches 28.4 dB PSNR compared with 29.0 dB for the sensor-rich reference. These results show that RGB-only geometry can recover much of the sensor-rich performance while producing one compact facade representation and explicitly flagging cases with insufficient cross-view support. The controlled experiments do not constitute an end-to-end city-scale GSV benchmark, but they establish the practical capability and trade-offs of sensor-light multi-view facade reconstruction.