• Stars
    star
    199
  • Rank 196,105 (Top 4 %)
  • Language
    Python
  • License
    GNU Affero Genera...
  • Created almost 3 years ago
  • Updated 8 months ago

Reviews

There are no reviews yet. Be the first to send feedback to the community and the maintainers!

Repository Details

[CVPR 2022] Official PyTorch Implementation for "Reference-based Video Super-Resolution Using Multi-Camera Video Triplets"

Reference-based Video Super-Resolution (RefVSR)
Official PyTorch Implementation of the CVPR 2022 Paper
Project | Paper | arXiv | RealMCVSR Dataset
Hugging Face Spaces License CC BY-NC
PWC

This repo contains training and evaluation code for the following paper:

Reference-based Video Super-Resolution Using Multi-Camera Video Triplets
Junyong Lee, Myeonghee Lee, Sunghyun Cho, and Seungyong Lee
POSTECH
IEEE Computer Vision and Pattern Recognition (CVPR) 2022


Getting Started

Prerequisites

Tested environment

Ubuntu Python PyTorch CUDA

1. Environment setup

$ git clone https://github.com/codeslake/RefVSR.git
$ cd RefVSR

$ conda create -y name RefVSR python 3.8 && conda activate RefVSR

# Install pytorch
$ conda install pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3 -c pytorch

# Install requirements
$ ./install/install_cudnn113.sh

It is recommended to install PyTorch >= 1.10.0 with CUDA11.3 for running small models using Pytorch AMP, because PyTorch < 1.10.0 is known to have a problem in running amp with torch.nn.functional.grid_sample() needed for inter-frame alignment.

For the other models, PyTorch 1.8.0 is verified. To install requirements with PyTorch 1.8.0, run ./install/install_cudnn102.sh for CUDA10.2 or ./install/install_cudnn111.sh for CUDA11.1

2. Dataset

Download and unzip the proposed RealMCVSR dataset under [DATA_OFFSET]:

[DATA_OFFSET]
    └── RealMCVSR
        β”œβ”€β”€ train                       # a training set
        β”‚   β”œβ”€β”€ HR                      # videos in original resolution 
        β”‚   β”‚   β”œβ”€β”€ T                   # telephoto videos
        β”‚   β”‚   β”‚   β”œβ”€β”€ 0002            # a video clip 
        β”‚   β”‚   β”‚   β”‚   β”œβ”€β”€ 0000.png    # a video frame
        β”‚   β”‚   β”‚   β”‚   └── ...         
        β”‚   β”‚   β”‚   └── ...            
        β”‚   β”‚   β”œβ”€β”€ UW                  # ultra-wide-angle videos
        β”‚   β”‚   └── W                   # wide-angle videos
        β”‚   β”œβ”€β”€ LRx2                    # 2x downsampled videos
        β”‚   └── LRx4                    # 4x downsampled videos
        β”œβ”€β”€ test                        # a testing set
        └── valid                       # a validation set

[DATA_OFFSET] can be modified with --data_offset option in the evaluation script.

3. Pre-trained models

Download pretrained weights (Google Drive | Dropbox) under ./ckpt/:

RefVSR
β”œβ”€β”€ ...
β”œβ”€β”€ ./ckpt
β”‚   β”œβ”€β”€ edvr.pytorch                    # weights of EDVR modules used for training Ours-IR
β”‚   β”œβ”€β”€ SPyNet.pytorch                  # weights of SpyNet used for inter-frame alignment
β”‚   β”œβ”€β”€ RefVSR_small_L1.pytorch         # weights of Ours-small-L1
β”‚   β”œβ”€β”€ RefVSR_small_MFID.pytorch       # weights of Ours-small
β”‚   β”œβ”€β”€ RefVSR_small_MFID_8K.pytorch    # weights of Ours-small-8K
β”‚   β”œβ”€β”€ RefVSR_L1.pytorch               # weights of Ours-L1
β”‚   β”œβ”€β”€ RefVSR_MFID.pytorch             # weights of Ours
β”‚   β”œβ”€β”€ RefVSR_MFID_8K.pytorch.pytorch  # weights of Ours-8K
β”‚   β”œβ”€β”€ RefVSR_IR_MFID.pytorch          # weights of Ours-IR
β”‚   └── RefVSR_IR_L1.pytorch            # weights of Ours-IR-L1
└── ...

For the testing and training of your own model, it is recommended to go through wiki pages for
logging and details of testing and training scripts before running the scripts.

Testing models of CVPR 2022

Evaluation script

CUDA_VISIBLE_DEVICES=0 python -B run.py \
    --mode _RefVSR_MFID_8K \                       # name of the model to evaluate
    --config config_RefVSR_MFID_8K \               # name of the configuration file in ./configs
    --data RealMCVSR \                             # name of the dataset
    --ckpt_abs_name ckpt/RefVSR_MFID_8K.pytorch \  # absolute path for the checkpoint
    --data_offset /data1/junyonglee \              # offset path for the dataset (e.g., [DATA_OFFSET]/RealMCVSR)
    --output_offset ./result                       # offset path for the outputs

Real-world 4x video super-resolution (HD to 8K resolution)

# Evaluating the model 'Ours' (Fig. 8 in the main paper).
$ ./scripts_eval/eval_RefVSR_MFID_8K.sh

# Evaluating the model 'Ours-small'.
$ ./scripts_eval/eval_amp_RefVSR_small_MFID_8K.sh

For the model Ours, we use Nvidia Quadro 8000 (48GB) in practice.

For the model Ours-small,

  • We use Nvidia GeForce RTX 3090 (24GB) in practice.
  • It is the model Ours-small in Table 2 further trained with the adaptation stage.
  • The model requires PyTorch >= 1.10.0 with CUDA 11.3 for using PyTorch AMP.

Quantitative evaluation (models trained with the pre-training stage)

## Table 2 in the main paper
# Ours
$ ./scripts_eval/eval_RefVSR_MFID.sh

# Ours-l1
$ ./scripts_eval/eval_RefVSR_L1.sh

# Ours-small
$ ./scripts_eval/eval_amp_RefVSR_small_MFID.sh

# Ours-small-l1
$ ./scripts_eval/eval_amp_RefVSR_small_L1.sh

# Ours-IR
$ ./scripts_eval/eval_RefVSR_IR_MFID.sh

# Ours-IR-l1
$ ./scripts_eval/eval_RefVSR_IR_L1.sh

For all models, we use Nvidia GeForce RTX 3090 (24GB) in practice.

To obtain quantitative results measured with the varying FoV ranges as shown in Table 3 of the main paper, modify the script and specify --eval_mode FOV.

Training models with the proposed two-stage training strategy

The pre-training stage (Sec. 4.1)

# To train the model 'Ours':
$ ./scripts_train/train_RefVSR_MFID.sh

# To train the model 'Ours-small':
$ ./scripts_train/train_amp_RefVSR_small_MFID.sh

For both models, we use Nvidia GeForce RTX 3090 (24GB) in practice.

Be sure to modify the script file and set proper GPU devices, number of GPUs, and batch size by modifying CUDA_VISIBLE_DEVICES, --nproc_per_node and -b options, respectively.

  • We use the total batch size of 4, the multiplication of numbers in options --nproc_per_node and -b.

The adaptation stage (Sec. 4.2)

  1. Set the path of the checkpoint of a model trained with the pre-training stage.
    For the model Ours-small, for example,

    $ vim ./scripts_train/train_amp_RefVSR_small_MFID_8K.sh
    #!/bin/bash
    
    py3clean ./
    CUDA_VISIBLE_DEVICES=0,1 ...
        ...
        -ra [LOG_OFFSET]/RefVSR_CVPR2022/amp_RefVSR_small_MFID/checkpoint/train/epoch/ckpt/amp_RefVSR_small_MFID_00xxx.pytorch
        ...
    

    Checkpoint path is [LOG_OFFSET]/RefVSR_CVPR2022/[mode]/checkpoint/train/epoch/[mode]_00xxx.pytorch.

    • PSNR is recorded in [LOG_OFFSET]/RefVSR_CVPR2022/[mode]/checkpoint/train/epoch/checkpoint.txt.
    • [LOG_OFFSET] can be modified with config.log_offset in ./configs/config.py.
    • [mode] is the name of the model assigned with --mode in the script used for the pre-training stage.
  2. Start the adaptation stage.

    # Training the model 'Ours'.
    $ ./scripts_train/train_RefVSR_MFID_8K.sh
    
    # Training the model 'Ours-small'.
    $ ./scripts_train/train_amp_RefVSR_small_MFID_8K.sh

    For the model Ours, we use Nvidia Quadro 8000 (48GB) in practice.

    For the model Ours-small, we use Nvidia GeForce RTX 3090 (24GB) in practice.

    Be sure to modify the script file to set proper GPU devices, number of GPUs, and batch size by modifying CUDA_VISIBLE_DEVICES, --nproc_per_node and -b options, respectively.

    • We use the total batch size of 2, the multiplication of numbers in options --nproc_per_node and -b.

Training models with L1 loss

# To train the model 'Ours-l1':
$ ./scripts_train/train_RefVSR_L1.sh

# To train the model 'Ours-small-l1':
$ ./scripts_train/train_amp_RefVSR_small_L1.sh

# To train the model 'Ours-IR-l1':
$ ./scripts_train/train_amp_RefVSR_small_L1.sh

For all models, we use Nvidia GeForce RTX 3090 (24GB) in practice.

Be sure to modify the script file and set proper GPU devices, number of GPUs, and batch size by modifying CUDA_VISIBLE_DEVICES, --nproc_per_node and -b options, respectively.

  • We use the total batch size of 8, the multiplication of numbers in options --nproc_per_node and -b.

Wiki

Contact

Open an issue for any inquiries. You may also have contact with [email protected]

License

License CC BY-NC

This software is being made available under the terms in the LICENSE file. Any exemptions to these terms require a license from the Pohang University of Science and Technology.

Acknowledgment

We thank the authors of BasicVSR and DCSR for sharing their code.

BibTeX

@InProceedings{Lee2022RefVSR,
    author    = {Junyong Lee and Myeonghee Lee and Sunghyun Cho and Seungyong Lee},
    title     = {Reference-based Video Super-Resolution Using Multi-Camera Video Triplets},
    booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
    year      = {2022}
}