Enhanced spatial-temporal freedom for video frame interpolation.

With the introduction of deformable convolution, the kernel-based Video Frame Interpolation (VFI) has made significant progress. However, there are still some problems, such as the limited spatial-temporal freedom degree in the sampling point extraction stage and the insufficient utilization of spatial-temporal information in the feature extraction stage of the kernel-based methods. In this paper, we propose a video frame interpolation method based on Enhanced Spatial-Temporal Freedom (ESTF), which consists of four modules for VFI. Specifically, we first combine 3D and deformable convolutions to extract spatial-temporal feature information in the proposed enhanced spatial-temporal feature extraction module. Then, we utilize an enhanced freedom fusion module to adaptively estimate parameters and generate intermediate frames by adaptive estimators and deformable fusion layers, respectively. Finally, a context extraction module and a residual contextual refinement module are utilized to extract context features for optimizing the generated frames. Extensive experimental results on various popular benchmarks such as Vimeo90K, GOPRO, and Adobe240 demonstrate that the proposed method achieves competitive performance against most existing methods, especially when dealing with complex motions. [ABSTRACT FROM AUTHOR]