Transferring knowledge from monocular completion for self-supervised monocular depth estimation

Monocular depth estimation is a very challenging task in computer vision, with the goal to predict per-pixel depth from a single RGB image. Supervised learning methods require large amounts of depth measurement data, which are time-consuming and expensive to obtain. Self-supervised methods are showing great promise, exploiting geometry to provide supervision signals through image warping. Moreover, several works leverage on other visual tasks (e.g. stereo matching and semantic segmentation) to further advance self-supervised monocular depth estimation. In this paper, we propose a novel framework utilizing monocular depth completion as an auxiliary task to assist monocular depth estimation. In particular, a knowledge transfer strategy is employed to enable monocular depth estimation to benefit from the effective feature representations learned by monocular depth completion task. The correlation between monocular depth completion and monocular depth estimation could be fully and effectively utilized in this framework. Only unlabeled stereo images are used in the proposed framework, which achieves a self-supervised learning paradigm. Experimental results on publicly available dataset prove that the proposed approach achieves superior performance to state-of-the-art self-supervised methods and comparable performance with supervised methods.