Self-supervised monocular depth learning from unknown cameras: Leveraging the power of raw data

Self-supervised monocular depth estimation from wild videos with unknown camera intrinsics is a practical and challenging task in computer vision. Most of the existing methods in literature employed a camera decoder and a pose decoder to estimate camera intrinsics and poses respectively, however, their performances would be degraded significantly in many complex scenarios with severe noise and large camera rotations. To address this problem, we propose a novel self-supervised monocular depth estimation method, which could be trained from wild videos with a joint optimization strategy for simultaneously estimating camera intrinsics and poses. In the proposed method, a depth encoder is employed to learn scene depth features, and then by taking these features as inputs, a Neighborhood Influence Module (NIM) is designed for predicting each pixel’s depth by fusing the depths of its neighboring pixels, which could explicitly enforce the depth accuracy. In addition, a knowledge distillation mechanism is introduced to learn a lightweight depth encoder from a large-scale depth encoder, for achieving a balance between computational speed and accuracy. Experimental results on four public datasets demonstrate that the proposed method outperforms some state-of-the-art methods in most cases. Moreover, once the proposed method is trained with a mixed set of different datasets, its performance would be further boosted in comparison to the proposed method trained with each involved single dataset. Codes are available at: https://github.com/ZhuYongChaoUSST/IntrLessMonoDepth.