Adaptive Self-supervised Depth Estimation in Monocular Videos

In this work, we develop and evaluate two adaptive strategies to self-supervised depth estimation methods based on view reconstruction. First, we propose an adaptive consistency loss that extends the usage of minimum re-projection to enforce consistency on the pixel intensities, structure, and feature maps. Moreover, we evaluate two approaches to use uncertainty to weigh the error contribution in the input frames. Finally, we improve our model with a composite visibility mask. The results show that the adaptive consistency loss can effectively combine photometric, structure and feature consistency terms. Moreover, weighting the error contribution using uncertainty can improve the performance of a simpler version of the model, but cannot improve them model when all improvements are considered. Finally, our combined model achieves competitive results when compared to state-of-the-art methods.