Dynamic Spatial Propagation Network for Depth Completion

Image-guided depth completion aims to generate dense maps with sparse measurements and corresponding RGB images. Currently, spatial propagation networks (SPNs) are the most popular affinity-based methods in completion, but they still suffer from representation limitation of fixed affinity over smoothing during iterations. Our solution is estimate independent matrices each SPN iteration, it over-parameterized heavy calculation.This paper introduces an efficient model that learns among neighboring pixels attention-based, dynamic approach. Specifically, Dynamic Spatial Propagation Network (DySPN) we proposed makes use a non-linear (NLPM). It decouples neighborhood into parts regarding different distances recursively generates attention refine these adaptive matrices. Furthermore, adopt diffusion suppression (DS) operation so converges at early stage prevent over-smoothing depth. Finally, order decrease computational cost required, also introduce three variations reduce amount neighbors attentions needed while retaining similar accuracy. In practice, our method requires less iteration match performance other SPNs yields better results overall. DySPN outperforms state-of-the-art (SoTA) on KITTI Depth Completion (DC) evaluation by time submission able yield SoTA NYU v2 dataset as well.