Exploit Domain-Robust Optical Flow in Domain Adaptive Video Semantic Segmentation

Domain adaptive semantic segmentation aims to exploit the pixel-level annotated samples on source domain assist of unlabeled target domain. For such a task, key is construct reliable supervision signals However, existing methods can only provide unreliable constructed by model (SegNet) that are generally domain-sensitive. In this work, we try find domain-robust clue more signals. Particularly, experimentally observe domain-robustness optical flow in video tasks as it mainly represents motion characteristics scenes. cannot be directly used since both them essentially represent different information. To tackle issue, first propose novel Segmentation-to-Flow Module (SFM) converts maps flows, named segmentation-based (SF), and then Segmentation-based Flow Consistency (SFC) method impose consistency between SF flow, which implicitly supervise training model. The extensive experiments two challenging benchmarks demonstrate effectiveness our method, outperforms previous state-of-the-art with considerable performance improvement. Our code available at https://github.com/EdenHazardan/SFC.