Abstract: Video traffic has gradually occupied the majority of mobile traffic, and video damage in unstable transmission remains a common and urgent problem. The difficulty of inpainting these damaged videos is that the holes randomly appear in random video frames, which are hard to be well settled with both low latency and high accuracy. We are the pioneer to look into the video inpainting task in unstable transmission and propose a low-latency video inpainting method which consists of two stages: In the coarsely inpainting stage, we achieve the extraction of damaged two-dimensional optical flow from reference frames, and establish a linear prediction model to coarsely inpaint the damaged frames according to the temporal consistency of motions. In the fine inpainting stage, a Partial Convolutional Frame Completion network(PCFC-Net) is proposed to synthesize all reference information and calculate a fine inpainting result. Compared with that of the state-of-the-art baselines, the waiting time for reference frames is greatly reduced while PSNR and SSIM are improved by 4.0%～12.7% on DAVIS dataset.

Key words: video inpainting, unstable transmission, partial CNN, linear prediction