Bridging distant space-time interactions is important for high-quality video inpainting with large moving masks. Most existing technologies exploit patch similarities within the frames, or leaverage large-scale training data to fill the hole along spatial and temporal dimensions. Recent works introduce promissing Transformer architecture into deep video inpainting to escape from the dominanace of nearby interactions and achieve superior performance than their baselines. However, such methods still struggle to complete larger holes containing complicated scenes. To alleviate this issue, we first employ a fast Fourier convolutions, which cover the frame-wide receptive field, for token representation. Then, the token passes through the seperated spatio-temporal transformer to explicitly moel the long-range context relations and simultaneously complete the missing regions in all input frames. By formulating video inpainting as a directionless sequence-to-sequence prediction task, our model fills visually consistent content, even under conditions such as large missing areas or complex geometries. Furthermore, our spatio-temporal transformer iteratively fills the hole from the boundary enabling it to exploit rich contextual information. We validate the superiority of the proposed model by using standard stationary masks and more realistic moving object masks. Both qualitative and quantitative results show that our model compares favorably against the state-of-the-art algorithms.
Full Text Finder