DS-Depth: Dynamic and Static Depth Estimation via a Fusion Cost Volume

Self-supervised monocular depth estimation methods typically rely on the reprojection error to capture geometric relationships between successive frames in static environments. However, this assumption does not hold dynamic objects scenarios, leading errors during view synthesis stage, such as feature mismatch and occlusion, which can significantly reduce accuracy of generated maps. To address problem, we propose a novel cost volume that exploits residual optical flow describe moving objects, improving incorrectly occluded regions volumes used previous work. Nevertheless, inevitably generates extra occlusions noise, thus alleviate by designing fusion module makes compensate for each other. In other words, occlusion from is refined volume, incorrect information eliminated volume. Furthermore, pyramid distillation loss photometric inaccuracy at low resolutions an adaptive direction large gradient regions. We conducted extensive experiments KITTI Cityscapes datasets, results demonstrate our model outperforms previously published baselines self-supervised estimation.