CVPR 2023 – A Dynamic Multi-Scale Voxel Flow Network for Video Prediction

In this episode we discuss A Dynamic Multi-Scale Voxel Flow Network for Video Prediction
by Xiaotao Hu, Zhewei Huang, Ailin Huang, Jun Xu, Shuchang Zhou. The paper proposes a Dynamic Multi-scale Voxel Flow Network (DMVFN) for video prediction using only RGB images. The proposed network is efficient and achieves better performance than previous methods that require extra inputs for promising performance. The core of DMVFN is a differentiable routing module that effectively perceives the motion scales of video frames and selects adaptive sub-networks for different inputs at the inference stage. DMVFN outperforms state-of-the-art iterative-based OPT on generated image quality and is an order of magnitude faster than Deep Voxel Flow.