DesNet: Decomposed Scale-Consistent Network for Unsupervised Depth Completion

Unsupervised depth completion aims to recover dense from the sparse one without using ground-truth annotation. Although measurement obtained LiDAR is usually sparse, it contains valid and real distance information, i.e., scale-consistent absolute values. Meanwhile, scale-agnostic counterparts seek estimate relative have achieved impressive performance. To leverage both inherent characteristics, we thus suggest model upon unsupervised frameworks. Specifically, propose decomposed learning (DSCL) strategy, which disintegrates into prediction global scale estimation, contributing individual benefits. But unfortunately, most existing frameworks heavily suffer holes due extremely input weak supervisory signal. tackle this issue, introduce guidance (GDG) module, attentively propagates reference target via novel dense-to-sparse attention. Extensive experiments show superiority of our method on outdoor KITTI, ranking 1st outperforming best KBNet more than 12% in RMSE. Additionally, approach achieves state-of-the-art performance indoor NYUv2 benchmark as well.