Unsupervised Hierarchical Domain Adaptation for Adverse Weather Optical Flow

Optical flow estimation has made great progress, but usually suffers from degradation under adverse weather. Although semi/full-supervised methods have good attempts, the domain shift between synthetic and real weather images would deteriorate their performance. To alleviate this issue, our start point is to unsupervisedly transfer knowledge source clean target degraded domain. Our key insight that does not change intrinsic optical of scene, causes a significant difference for warp error images. In work, we propose first unsupervised framework via hierarchical motion-boundary adaptation. Specifically, employ image translation construct transformation relationship domains. motion adaptation, utilize consistency align cross-domain flows into motion-invariance common space, where used as guidance-knowledge obtain preliminary Furthermore, leverage inconsistency which measures misalignment boundary domains, joint intra- inter-scene contrastive adaptation refine boundary. The jointly promotes in unified framework. Extensive quantitative qualitative experiments been performed verify superiority proposed method.