Video Frame Interpolation via Adaptive Separable Convolution

TLDR

Standard video frame interpolation first estimates optical flow and then synthesizes intermediate frames, while recent methods merge these steps into a single convolution with spatially adaptive kernels; however, the large kernels needed for substantial motion limit pixel coverage due to memory constraints. This paper proposes formulating frame interpolation as local separable convolution over input frames using pairs of 1D kernels to reduce kernel size. The authors develop a deep fully convolutional neural network that takes two input frames, estimates per‑pixel pairs of 1D kernels, and synthesizes the entire intermediate frame end‑to‑end, trained on unannotated video data with perceptual loss. Experiments demonstrate that the 1D separable convolution approach requires far fewer parameters and produces high‑quality interpolated frames, outperforming conventional 2D kernel methods in both qualitative and quantitative evaluations.

Abstract

Standard video frame interpolation methods first estimate optical flow between input frames and then synthesize an intermediate frame guided by motion. Recent approaches merge these two steps into a single convolution process by convolving input frames with spatially adaptive kernels that account for motion and re-sampling simultaneously. These methods require large kernels to handle large motion, which limits the number of pixels whose kernels can be estimated at once due to the large memory demand. To address this problem, this paper formulates frame interpolation as local separable convolution over input frames using pairs of 1D kernels. Compared to regular 2D kernels, the 1D kernels require significantly fewer parameters to be estimated. Our method develops a deep fully convolutional neural network that takes two input frames and estimates pairs of 1D kernels for all pixels simultaneously. Since our method is able to estimate kernels and synthesizes the whole video frame at once, it allows for the incorporation of perceptual loss to train the neural network to produce visually pleasing frames. This deep neural network is trained end-to-end using widely available video data without any human annotation. Both qualitative and quantitative experiments show that our method provides a practical solution to high-quality video frame interpolation.