Optical Flow Distillation: Towards Efficient and Stable Video Style Transfer

Abstract

Video style transfer techniques inspire many exciting applications on mobile devices. However, their efficiency and stability are still far from satisfactory. To boost the transfer stability across frames, optical flow is widely adopted, despite its high computational complexity, e.g. occupying over 97% inference time. This paper proposes to learn a lightweight video style transfer network via knowledge distillation paradigm. We adopt two teacher networks, one of which takes optical flow during inference while the other does not. The output difference between these two teacher networks highlights the improvements made by optical flow, which is then adopted to distill the target student network. Furthermore, a low-rank distillation loss is employed to stabilize the output of student network by mimicking the rank of input videos. Extensive experiments demonstrate that our student network without an optical flow module is still able to generate stable video and runs much faster than the teacher network.

Style-Aware Normalized Loss for Improving Arbitrary Style Transfer

“…Finally, experimental results show up to 111% and 98% relative improvements in Deception Rate and human preference, respectively. Future work can adopt our new loss in related problems, e.g., video [1,3,5,12,18,29,40,41] and photo [29,35,38,53] stylization, texture synthesis [10,11,26,51], etc. Future work can also derive tighter bounds for the style loss to improve style-aware normalization.…”

Learning to Stylize Novel Views

“…Video stylization aims to transfer the style of a reference image to a sequence of video frames. To address the temporal flickering issue produced by the image stylization approaches, numerous approaches [3,5,8,12,14] incorpo-rate optical flow modules to train feed-forward networks for transferring a particular style to the videos. Several recent frameworks [7,9,52] enable the video style transfer to arbitrary styles.…”

Learning Long Term Style Preserving Blind Video Temporal Consistency

“…Following up on [12] we formulate the hypothesis on stable stylized/processed videos: non-occluded regions should be low-rank representations. Consider during training (i) a sequence of k consecutive frames, {I t , .…”

“…Based on the formulated hypothesis, the rank of χ constructed using the raw input frames I t , χ I and using the output frames of the model O t , χ O should not be too different from each other. [12] propose the low rank loss using the convex relaxation of the rank defined as…”