Seeing Farther Than Supervision: Self-supervised Depth Completion in Challenging Environments

This paper tackles the problem of learning a depth completion network from a series of RGB images and short-range depth measurements as a new setting for depth completion. Commodity RGB-D sensors used in indoor environments can provide dense depth measurements; however, their acquisition distance is limited. Recent depth completion methods train CNNs to estimate dense depth maps in a supervised/self-supervised manner while utilizing sparse depth measurements. For self-supervised learning, indoor environments are challenging due to many non-textured regions, leading to the problem of inconsistency. To overcome this problem, we propose a self-supervised depth completion method that utilizes optical flow from two RGB-D images. Because optical flow provides accurate and robust correspondences, the ego-motion can be estimated stably, which can reduce the difficulty of depth completion learning in indoor environments. Experimental results show that the proposed method outperforms the previous self-supervised method in the new depth completion setting and produces qualitatively adequate estimates.