Skip to content

Latest commit

 

History

History
121 lines (99 loc) · 5.58 KB

README.md

File metadata and controls

121 lines (99 loc) · 5.58 KB
Raw

Video Harmonization with Triplet Spatio-Temporal Variation Patterns

Here we provide the PyTorch implementation and pre-trained model of our latest version.

Prerequisites

  • Linux
  • Python 3
  • CPU or NVIDIA GPU + CUDA CuDNN

Video Harmonization

Train/Test

  • Download HYouTube dataset.

  • Train our VTT model:

CUDA_VISIBLE_DEVICES=0 python train.py --model basestv2 --netG vth --name experiment_name --deformable_depth x --gpt_depth x --dataset_root <dataset_dir> --batch_size x --init_port xxxx --loss_T --save_iter_model
  • Test our VTT model:
CUDA_VISIBLE_DEVICES=0 python test.py --model basestv2 --netG  vth  --name experiment_name --deformable_depth x --gpt_depth x --dataset_root <dataset_dir> --batch_size 1  --init_port xxxx

Apply a pre-trained model

  • Download pre-trained models from BaiduCloud (access code: h63l), and put latest_net_G.pth in the directory checkpoints/vth_harmonization. Run:
CUDA_VISIBLE_DEVICES=0 python test.py --model basestv2 --netG  vth  --name vth_harmonization --deformable_depth 2 --gpt_depth 2 --dataset_root <dataset_dir> --batch_size 1  --init_port xxxx

Evaluation

To evaluate the spatial consistency, run:

CUDA_VISIBLE_DEVICES=0 python evaluation/ih_evaluation.py --dataroot <dataset_dir> --result_root results/experiment_name/test_latest/images/ --evaluation_type hyt --dataset_name HYT

To evaluate the temporal consistency, run:

python tc_evaluation.py --dataset_root <dataset_dir> --experiment_name experiment_name --mode 'v' --brightness_region 'foreground'

Real composite image harmonization

More compared results can be found at BaduCloud (access code: h63l).

Video Enhancement

Train/Test

  • Download SDSD dataset.

  • Train our VTT model:

CUDA_VISIBLE_DEVICES=0 python train.py --model basestv2 --netG vth --name experiment_name   --deformable_depth x --gpt_depth x  --dataset_root <dataset_dir> --batch_size x --init_port xxxx --n_frames 5 --loss_T --save_iter_model
  • Test our VTT model:
CUDA_VISIBLE_DEVICES=1 python test.py --model basestv2 --netG vth --name experiment_name --deformable_depth x --gpt_depth x --dataset_root <dataset_dir> --batch_size 1 --init_port xxxx --n_frames 15

Apply a pre-trained model

  • Download pre-trained models from BaiduCloud (access code: h63l), and put latest_net_G.pth in the directory checkpoints/vth_enhancement. Run:
CUDA_VISIBLE_DEVICES=0 python test.py --model basestv2 --netG vth --name vth_enhancement --deformable_depth 2 --gpt_depth 2 --dataset_root <dataset_dir> --batch_size 1 --init_port xxxx --n_frames 15

Evaluation

To evaluate the spatial consistency, run:

CUDA_VISIBLE_DEVICES=0 python evaluation/ve_evaluation.py --dataroot <dataset_dir> --result_root results/experiment_name/test_latest/images/input/

To evaluate the temporal consistency, run:

python tc_evaluation.py --dataset_root <dataset_dir> --experiment_name experiment_name --mode 'v' --brightness_region 'image'

Video Demoireing

Train/Test

CUDA_VISIBLE_DEVICES=0 python train.py --model basestv2 --netG  vth  --name experiment_name --deformable_depth x --gpt_depth x --dataset_root <dataset_dir> --batch_size x  --init_port xxxx --n_frames 5  --loss_T   --save_iter_model
  • Test our VTT model:
CUDA_VISIBLE_DEVICES=1 python test.py --model basestv2 --netG  vth  --name experiment_name --deformable_depth x --gpt_depth x --dataset_root <dataset_dir>  --batch_size 1 --init_port xxxx --n_frames 20

Apply a pre-trained model

  • Download pre-trained models from BaiduCloud (access code: h63l), and put latest_net_G.pth in the directory checkpoints/vth_demoireing. Run:
CUDA_VISIBLE_DEVICES=0 python test.py --model basestv2 --netG  vth  --name vth_demoireing --deformable_depth 2 --gpt_depth 2 --dataset_root <dataset_dir> --batch_size 1 --init_port xxxx --n_frames 20

Evaluation

To evaluate the spatial consistency, run:

CUDA_VISIBLE_DEVICES=0 python evaluation/vd_evaluation.py --dataroot <dataset_dir> --result_root results/experiment_name/test_latest/images/test/source/

To evaluate the temporal consistency, run:

python tc_evaluation.py --dataset_root <dataset_dir> --experiment_name experiment_name --mode 'v' --brightness_region 'image'

Acknowledgement

For some of the data modules and model functions used in this source code, we need to acknowledge the repositories of HarmonyTransformer, Swin3D, VEN-Retinex and VDRTC.