Abstract
We study video inpainting, which aims to recover realistic textures from damaged frames. Recent progress has been made by taking other frames as references so that relevant textures can be transferred to damaged frames. However, existing video inpainting approaches neglect the ability of the model to extract information and reconstruct the content, resulting in the inability to reconstruct the textures that should be transferred accurately. In this paper, we propose a novel and effective spatial-temporal texture transformer network (STTTN) for video inpainting. STTTN consists of six closely related modules optimized for video inpainting tasks: feature similarity measure for more accurate frame pre-repair, an encoder with strong information extraction ability, embedding module for finding a correlation, coarse low-frequency feature transfer, refinement high-frequency feature transfer, and decoder with accurate content reconstruction ability. Such a design encourages joint feature learning across the input and reference frames. To demonstrate the advancedness and effectiveness of the proposed model, we conduct comprehensive ablation learning and qualitative and quantitative experiments on multiple datasets by using standard stationary masks and more realistic moving object masks. The excellent experimental results demonstrate the authenticity and reliability of the STTTN.