The PIRM-SR Challenge will compare and rank methods for perceptual single-image super-resolution. State-of-the-art methods in terms of perceptual quality (e.g. SRGAN) are rated poorly by "simple" distortion measures such as PSNR and SSIM. Therefore, in contrast to previous challenges, the evaluation and ranking will be done in a perceptual-quality aware manner based on [Blau and Michaeli, CVPR'18]. This unified approach quantifies the accuracy and perceptual quality of algorithms jointly, and will enable perceptual-driven methods to compete alongside algorithms that target PSNR maximization.
For further details see the challenge website. The PIRM dataset can be found in this link
This Matlab code computes the RMSE and perceptual scores for your method's outputs on the self-validation set.
- Copy your outputs into the
your_resultsfolder in the base directory. - Copy the validation/test set (HR images only) into the
self_validation_HRfolder. - Download the Ma et al. code, and extract it into the
utils/sr-metric-masterfolder. - Run the
evaluate_results.mscript.
Depending on your operating system, you may need the recompile the MEX files in the matlabPyrTools toolbox. If so:
- Run
utils/sr-metric-master/external/matlabPyrTools/MEX/compilePyrTools.m - Copy the generated MEX files into the parent directory
utils/sr-metric-master/external/matlabPyrTools
Note: in Linux or OS you should also change line 82 in mex_regressionRF_predict.cpp to: plhs[0]=mxCreateNumericMatrix(n_size,1,mxDOUBLE_CLASS,mxREAL);
Pre-compiled mex files (for OS, Linux and Win) are also available at this link (Thank you Muhammad Haris for the solution).
This code is distributed only for academic research purposes only.
For other purposes, please contact Roey Mechrez: roey (at) campus.technion.ac.il