CVPR2023-DMVFN

DMVFN is an open-source PyTorch implementation of the Dynamic Multi-Scale Voxel Flow Network for video frame prediction, presented at CVPR 2023. It predicts future video frames by analyzing motion and visual information from preceding frames. The architecture uses dynamic multi-scale voxel flow to estimate motion across different spatial resolutions. This approach helps model complex movements while managing computational requirements. The repository includes training and evaluation workflows for Cityscapes and KITTI driving datasets. Pretrained model weights are provided for reproducing experiments and running predictions. Researchers can explore inference examples through a Google Colab notebook. DMVFN is intended for computer vision research involving video forecasting, motion estimation, and temporal scene understanding.