Skip to content

Latest commit

 

History

History
39 lines (34 loc) · 2.06 KB

README.md

File metadata and controls

39 lines (34 loc) · 2.06 KB

This is the official repo for NeurIPS 2024 paper - Motion Graph Unleashed: A Novel Approach to Video Prediction [paper][poster]

TODO

  • code base
  • Data preparation
  • UCF Sports STRPM Dataloader
  • Download link to all testing results images

Prepare

git clone https://github.com/Kay1794/Motion-Graph-Video-Prediction.git
cd Motion-Graph-Video-Prediction
conda env create -f environment.yml

Dataset

UCF Sports Dataset

  • Data download
  • MMVP split
    • Unzip downloaded file
    • Modify ucf_config['dataroot'] in config.py to the unziped folder
    • Copy split txt files to ucf_config['dataroot']
  • STRPM split
    • Dataloader
    • Implement steps

KITTI & Cityscapse

Results download [Google Drive]

Run

Training on UCF Sports MMVP Split

python main.py --mode train --scale_in_use 4 --base_channel 16 --downsample_scale 2 2 2 --exp baseline --cos_restart --rot_aug --flip_aug --loss_list recon --edge_normalize --pred_att_iter_num 3 --tendency_len 16 --edge_list backward forward spatial --t_period 300 --nepoch 300 --eval_list psnr ssim lpips --logpath ./results/ --shuffle_scale 2 --pos_len 4 --loss_list recon --top_k 0.01 --batch 16 --dataset ucf_4to1 --energy_save_mode --log

Validation on UCF Sports MMVP Split

python main.py --mode train  --scale_in_use 4 --base_channel 16  --downsample_scale 2 2 2 --exp baseline --cos_restart --rot_aug --flip_aug --loss_list recon --edge_normalize --pred_att_iter_num 3 --tendency_len 16 --edge_list backward forward spatial  --t_period 300 --nepoch 300 --eval_list psnr ssim lpips --shuffle_scale 2 --pos_len 4 --loss_list recon --top_k 0.01 --batch 16 --dataset ucf_4to1 --resume ./pretrained_model/ucf_mmvp_split.pth --mode val