Semantic-Guided Completion Network for Video Inpainting in Complex Urban Scene

Video inpainting aims to fill damaged areas in video frames with appropriate content. Complex scene contain cluttered or ambiguous semantics and objects, making video inpainting in such scenarios a challenging yet meaningful task. Current methods are limited by the lack of sufficient video information, resulting in blurred results and temporal artifacts. In this paper, we design a novel semantic-guided completion network that uses the semantic information of the videos to complete missing regions in the complex urban scene. Specifically, we first leverage the semantic information to model the structure and content of the video and improve U-Net network to complete the broken semantic image. Then, we propose a module based on spatial-adaptive normalization to guide the generation of the damaged part of the video pixels by combining semantic information. Our model's ability to generate reasonable and accurate content is demonstrated through both quantitative and qualitative results on two publicly available urban scene datasets.