Abstract

Depth measurement serves as a pivotal technology with widespread applications, but limitations of commodity depth sensors often result in noisy and incomplete depth maps for indoor scenes. In response, this study introduces a discrete geometric processing-based method aimed at enhancing the completeness of indoor depth maps, particularly in intricate and challenging scenes. The proposed approach involves transforming raw depth maps into a quadrilateral mesh surface and addressing missing depth information through a mesh deformation framework. The mesh deformation framework employs iterations to update constraints on quadrilateral facets until convergence, tackling normal constraints, position constraints, and perspective projection constraints. This optimization problem is resolved using a combination of local shaping and global fusing strategies. In the local step, each quadrilateral facet is projected from the physical imaging plane into camera coordinate system based on perspective projection, preserving only orientation information. The global step then propagates location information through vertices shared by connected facets. Uncertainties in input normal and observed depth information are addressed through updating strategies. Experimental results on 30 selected examples from ScanNet dataset demonstrate the superiority of the proposed method over two existing techniques quantitatively and qualitatively. Specifically, the method excels in addressing depth discontinuity around object boundaries, producing clear disconnections, while competing methods exhibit errors and stretched distortions. Validating normal and depth updating strategies, the paper confirms the effectiveness of processing depth discontinuity in reducing completion errors. The robustness of the method is further affirmed by energy reductions in both normal and position aspects across iterations.