DESC: Domain Adaptation for Depth Estimation via Semantic Consistency

Abstract Accurate real depth annotations are difficult to acquire, needing the use of special devices such as a LiDAR sensor. Self-supervised methods try overcome this problem by processing video or stereo sequences, which may not always be available. Instead, in paper, we propose domain adaptation approach train monocular estimation model using fully-annotated source dataset and non-annotated target dataset. We bridge gap leveraging semantic predictions low-level edge features provide guidance for domain. enforce consistency between main second trained with segmentation maps, introduce priors form instance heights. Our is evaluated on standard benchmarks show consistent improvement upon state-of-the-art. Code available at https://github.com/alopezgit/DESC .