Let the Monocular Depth Estimation Model See Further

Deep Neural Networks (DNN) based Monocular Depth Estimation (MDE) methods have recently made remarkable progress. However, we observe that some self-supervised and supervised methods suffer from the Depth-Sensitive Error Expansion (DSEE) issue-depth estimation error enlarges as the depth of the target scene increases. In particular, for selfsupervised methods, the gradient norm of the photometric error w.r.t. the estimated depth is inversely proportional to the square of true depth, which implies the larger the depth is, the smaller the gradient norm it produces and thus the more difficult it to update the model. A similar phenomenon also exists in supervised models. We propose an effective fusion method to address this issue. Our method combines the best of current DNN based methods and Semi-Global Matching algorithm (SGM). Extensive experiments show that our fusion strategy can boost a variety of both self-supervised and supervised approaches.