Gaussian Surface Reconstruction with Adaptive Depth Supervision for Geographic Information Systems
Year: 2026
Type: preprint
Abstract: Surface reconstruction from overlapping images supplies geometry for three-dimensional geographic information systems (GIS), but stereo depth errors can constrain the recovered surfaces. We present HelixGau, a Gaussian surface reconstruction method that filters and updates stereo depth during optimization. Local correlation with monocular depth selects pixels for supervision, while stereo estimates supply the supervised values. Rendered planes with higher photometric correlation across views provide candidates for stereo plane updates. The updated depth maps are filtered before supervision, and surfaces are extracted after optimization. On 16 BlendedMVS scenes, mean F1 is 88.64 with a standard deviation of 0.09 over four runs, compared with the highest baseline mean of 76.85. Mean F1 is 0.500 on six Tanks and Temples scenes. On the Technical University of Denmark (DTU) benchmark, overall error is 0.505 mm, compared with the lowest baseline error of 0.504 mm. Component ablations measure the separate effects of filtering and updates on the reconstructed geometry.
Source: Research Square
Institutions: Wuhan University, Chinese Academy of Sciences
Language: English
Topic: Advanced Vision and Imaging
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7234071470
