Fast Asymmetric Bilateral Motion Estimation for Video Frame Interpolation

We propose a fast asymmetric bilateral motion estimation (f-ABME) network for video frame interpolation, which is based on fast flow estimation networks. First, we estimate symmetric bilateral motion fields, assuming that motion trajectories between two input frames are linear. Second, we use the symmetric fields to generate an anchor frame. Third, we estimate asymmetric bilateral motion fields from the anchor frame to the input frames. Finally, we use both the symmetric and asymmetric fields to reconstruct the intermediate frame. Experimental results demonstrate that the proposed f-ABME algorithm achieves competitive VFI results, while demanding relatively low complexity.