Multi-scale Spatial Propagation Network for Depth Completion

Depth Completion focuses on creating an accurate, dense, pixel-wise depth from sparse depth. Prior approaches utilized RGB images as guides and applied spatial propagation to refine initial dense depth maps. However, these methods are constrained by limited receptive fields. They require multiple iterations for full image depth propagation. The large number of parameters and iterations hinders practical applications. To over-come these limitations, this paper presents a Multi-scale Convolutional Spatial Propagation Network. This network effectively broadens the iterative refinement's receptive field and introduces an adaptive neighborhood propagation. Firstly, it performs iterative refinement on multi-scaled depth maps, scaling the receptive field accordingly. Secondly, we introduce an innovative neighbor-hood selection strategy. This strategy reduces the computational complexity associated with expanding neighborhoods. Addition-ally, this method introduces direction-aware affinity normalization. During the propagation phase, it adaptively adjusts affinity. This preserves object boundary sharpness. Extensive tests conducted on the NYU Depth v2 dataset confirm the effectiveness of our approach. Compared to other SPN-based methods, our approach achieves efficient neighborhood propagation, and requires only three iterations to produce detailed dense depth maps.