We propose a method for acceleration of stereo camera simulation using stereo super resolution (SSR). This repository mainly contains the implementation of our SSR model (ETSSR) and other super-resolution (SR) models cited in the paper.

You can view the examples of ETSSR super-resolved images in the provided demo.

• Installation
• Training
• Testing
• Profiling
• Acknowledgements

• Create a conda environment and activate it.

  $ conda create --name etssr python=3.6

  $ conda activate etssr

• Install PyTorch following official instructions

The code has been tested on Ubuntu 18.04 with CUDA 11.4 and Pytorch 1.10.2.

• Install other python packages listed in requirements.txt

  $ pip install -r requirements.txt

• Download our CMRSI dataset and unzip it.

• To train the super-resolution (SR) models (including ETSSR) on CMRSI dataset, first generate image patches of CMRSI using:

python generate_patches.py

You need to set the directory of downloaded CMRSI dataset and the destination directory of image patches in the generate_patches.py code.

• Set the training parameters in the train.yaml file located in the ./configs folder.

You must change the "data_dir" to the path to the CMRSI image patches and "test_data_dir" to the path to the orignal CMRSI dataset.

• Run the training code using:

python train.py --cfg ./configs/train.yaml

The log of the training, including tensorboard plots and the model weights are saved in "checkpoint_dir"/"exp_name".

• To test the SR models on the CMRSI test set, set the testing parameters in the test.yaml file located in the ./configs folder.

You must change "data_dir" parameter in the config file to the path to CMRSI dataset.

• Run the test code using:

python test.py --cfg ./configs/test.yaml

The code calculates PSNR/SSIM on the CMRSI test set and saves output images in the "checkpoint_dir".

To calculate the number of parameters and Flops of SR models run:

python profiler.py

Set the model name inside the python code.

• State-of-the-art SR models implementations from YingqianWang

• Swin-transformer Layers from JingyunLiang