Deep Sparse Depth Completion Using Multi-Affinity Matrix

Image-guided depth completion aims to generate dense maps from sparse guided by their corresponding color (RGB) images. In this paper, we propose deep using multi-affinity matrix. Recently, spatial propagation networks (SPNs) are used refine obtained initial completion. However, they use the same affinity matrix even in multiple iterations that has a limit improving performance, which is not effective considering relationship between adjacent pixels. Thus, replace it with represent an output pixel and its neighboring ones. Moreover, neural can effectively fuse features two different modalities based on confidence maps. Inspired dynamic filtering, convolutional network (CSPN) multi-modal at stages. When training branch of proposed network, adopt supervised learning constrains all layers decoder. Since spatially varying required for feature fusion, produces adaptive single Experimental results KITTI NYU-v2 datasets show represents dependency pixels predicts accurate values calculating weights The achieves state-of-the-art performance terms root mean square error (RMSE) absolute (MAE).