Geometry-Aware Reference Synthesis for Multi-View Image Super-Resolution

This is the official implementation of the paper Geometry-Aware Reference Synthesis for Multi-View Image Super-Resolution, ACM MM 2022.

Setup

We use NVIDIA RTX 3090, cuda 11.1 and the follwing python packages:

pytorch=1.9.0
torchvision=0.10.0
pillow=8.3.1

Our code borrows from FreeViewSynthesis, please build the Python extension needed for preprecessing,

cd ext/preprocess
cmake -DCMAKE_BUILD_TYPE=Release .
make

Run Geometry-Aware Reference Synthesis (GARS) module to synthesize references

We preprocess and reorganize the folder structure for three public datasets, Tanks and Temples, BlendedMVS, and GTAV, for training and evaluation. In addition, we provide the preprocessing code for Tanks and Temples datasets in preprocess_Tanks_Temples.py. You can download our preprocessed version from baidudisk, code: ey16, and place them in the GARS folder. Then run:

cd GARS

# Synthesize references for Tanks and Temples:
bash GARS_Tanks_Temples.sh  // The CPU processes the Courthouse, Palace, and Church scenes due to a large number of images.

# Synthesize references for BlendedMVS:
bash GARS_BlendedMVS.sh  

# Synthesize references for GTAV:
bash GARS_GTAV.sh

Finallly, we will get Tanks_and_Temples_HLBIC_FR_MVRs_NVRs, BlendedMVS_HLBIC_FR_MVRs_NVRs, and GTAV_720_HLBIC_FR_MVRs_NVRs floders in the GARS folder. You can directly download the GARS results from baidudisk, code: aq9t.

Train Dynamic High-Frequency Search (DHFS) network

You can prepare dataset and updapte config.py with your own path.

cd exp

# Training the GAN-Based model, set use_perceptual_gan_loss=True in config.py
CUDA_VISIBLE_DEVICES=0 python exp.py --cmd retrain

# Training the Pixel-Based model, set use_perceptual_gan_loss=False in config.py
CUDA_VISIBLE_DEVICES=0 python exp.py --cmd retrain

Test

You can download our pretrained GAN-based model baidudisk (ctrq) and Pixel-based model baidudisk (jha1), and place them in exp/experiments. Then run:

cd exp

# Test the GAN-Based model, set use_perceptual_gan_loss=True in config.py
# Test the Pixel-Based model, set use_perceptual_gan_loss=False in config.py

# for Tanks and Temples, set blendedmvs = False and gtav = False in config.py
CUDA_VISIBLE_DEVICES=0 python exp.py --cmd eval --iter 139999

# for BlendedMVS, set blendedmvs = True and gtav = False in config.py
CUDA_VISIBLE_DEVICES=0 python exp.py --cmd eval --iter 139999

# for GTAV, set blendedmvs = False and gtav = True in config.py
CUDA_VISIBLE_DEVICES=0 python exp.py --cmd eval --iter 129999

Metrics

We provide the code to calculate PSNR, SSIM and LPIPS. Remember to update your own path.

cd metric

# for Tanks and Temples:
CUDA_VISIBLE_DEVICES=0 python result_tanks_psnr_ssim_lpips.py

# for BlendedMVS:
CUDA_VISIBLE_DEVICES=0 python result_blendedmvs_psnr_ssim_lpips.py

# for GTAV:
CUDA_VISIBLE_DEVICES=0 python result_gtav_psnr_ssim_lpips.py

Citation

If you find this code useful, please cite our paper:

@misc{https://doi.org/10.48550/arxiv.2207.08601,
  title = {Geometry-Aware Reference Synthesis for Multi-View Image Super-Resolution},
  author = {Cheng, Ri and Sun, Yuqi and Yan, Bo and Tan, Weimin and Ma, Chenxi},
  publisher = {arXiv},
  year = {2022},
}

Credit

Our code borrows from FreeViewSynthesis, RCAN, and C2-Matching .

Orange066 / MVSR