DSF-Net: Dual-Stream Fused Network for Video Frame Interpolation

Video frame interpolation aims to improve the video quality by increasing the frame rate. Existing methods adopt the cascaded architecture. They first estimate intermediate flow maps and then refine the synthesized intermediate frames with contextual features separately. However, the separated flow estimation and refined module ignore the mutual facilitation of them in frame interpolation. Following this issue, we propose a Dual-Stream Fused Network (DSF-Net) to joint flow estimation and refinement module for frame interpolation. Specifically, it first extracts the contextual features from input frames by a contextual feature extractor module, and then jointly refines the intermediate flow maps with the extracted features through a coarse-to-fine frame synthesis module. DSF-Net allows the intermediate flow and the contextual features to benefit each other while generating sharper moving objects and capturing better textual details. Experimental results demonstrate that DSF-Net performs consistently better than existing SOTA methods.