Unsupervised Monocular Depth Estimation in Highly Complex Environments

With the development of computational intelligence algorithms, unsupervised monocular depth and pose estimation framework, which is driven by warped photometric consistency, has shown great performance in day-time scenario. While some challenging environments, like night rainy night, essential consistency hypothesis untenable because complex lighting reflection, so that above framework cannot be directly applied to these scenarios. In this paper, we investigate problem highly scenarios address adopting an image transfer-based domain adaptation framework. We adapt model trained on applicable night-time scenarios, constraints both feature space output promote learn key features for decoding. Meanwhile, further tackle effects unstable transfer quality adaptation, approach proposed evaluate transferred images re-weight corresponding losses, as improve adapted model. Extensive experiments show effectiveness estimating dense map from images.