ICDC: Guided Interpolation and Correction with Depth Estimation Prior for Depth Completion

A depth estimation method that integrates monocular relative depth prior and sparse guided interpolation correction is proposed to address the problem of depth sparsity or loss in scenarios such as autonomous driving and industrial quality inspection. This method takes RGB images and pixel aligned sparse depth maps as inputs. Firstly, the DepthAnythingV2 model is used to extract the relative depth map and multiscale features of the image, providing accurate object edge and position relationship priors for subsequent tasks; Secondly, a relative depth guided interpolation strategy based on spatial distance KNN neighborhood is designed, which achieves preliminary alignment of relative and absolute depth through global fractional linear transformation. Combined with the optimization of linear transformation parameters of sparse truth points and their neighborhoods, the initial generation of sparse depth to dense depth is completed, effectively preserving object boundary details; Finally, an RDResUnet correction network incorporating DepthAnythingV2 features is constructed. The interpolated dense depth map and relative depth map are used as inputs, and the neighborhood mutation problem is corrected through residual learning to output a high-precision absolute dense depth map. Experiments have shown that this method has good adaptability to various sparse input forms such as random scatter points and LiDAR point clouds. Under extreme conditions where the sparse point ratio is less than 0.05%, it can still restore a true and complete global dense depth map, showing excellent edge preservation and depth accuracy.