PyTorch Image Quality (PIQ) is not endorsed by Facebook, Inc.;
PyTorch, the PyTorch logo and any related marks are trademarks of Facebook, Inc.
PyTorch Image Quality (PIQ) is a collection of measures and metrics for image quality assessment. PIQ helps you to concentrate on your experiments without the boilerplate code. The library contains a set of measures and metrics that is continually getting extended. For measures/metrics that can be used as loss functions, corresponding PyTorch modules are implemented.
We provide:
- Unified interface, which is easy to use and extend.
- Written on pure PyTorch with bare minima of additional dependencies.
- Extensive user input validation. Your code will not crash in the middle of the training.
- Fast (GPU computations available) and reliable.
- Most metrics can be backpropagated for model optimization.
- Supports python 3.7-3.10.
PIQ was initially named PhotoSynthesis.Metrics.
PyTorch Image Quality (PIQ) can be installed using pip, conda or git.
If you use pip, you can install it with:
$ pip install piqIf you use conda, you can install it with:
$ conda install piq -c photosynthesis-team -c conda-forge -c PyTorchIf you want to use the latest features straight from the master, clone PIQ repo:
git clone https://github.com/photosynthesis-team/piq.git
cd piq
python setup.py installThe full documentation is available at https://piq.readthedocs.io.
The group of metrics (such as PSNR, SSIM, BRISQUE) takes an image or a pair of images as input to compute a distance between them. We have a functional interface, which returns a metric value, and a class interface, which allows to use any metric as a loss function.
import torch
from piq import ssim, SSIMLoss
x = torch.rand(4, 3, 256, 256, requires_grad=True)
y = torch.rand(4, 3, 256, 256)
ssim_index: torch.Tensor = ssim(x, y, data_range=1.)
loss = SSIMLoss(data_range=1.)
output: torch.Tensor = loss(x, y)
output.backward()For a full list of examples, see image metrics examples.
The group of metrics (such as IS, FID, KID) takes a list of image features to compute the distance between distributions.
Image features can be extracted by some feature extractor network separately or by using the compute_feats method of a
class.
- Note:
compute_featsconsumes a data loader of a predefined format.
import torch
from torch.utils.data import DataLoader
from piq import FID
first_dl, second_dl = DataLoader(), DataLoader()
fid_metric = FID()
first_feats = fid_metric.compute_feats(first_dl)
second_feats = fid_metric.compute_feats(second_dl)
fid: torch.Tensor = fid_metric(first_feats, second_feats)If you already have image features, use the class interface for score computation:
import torch
from piq import FID
x_feats = torch.rand(10000, 1024)
y_feats = torch.rand(10000, 1024)
msid_metric = MSID()
msid: torch.Tensor = msid_metric(x_feats, y_feats)For a full list of examples, see feature metrics examples.
| Acronym | Year | Metric |
|---|---|---|
| PSNR | - | Peak Signal-to-Noise Ratio |
| SSIM | 2003 | Structural Similarity |
| MS-SSIM | 2004 | Multi-Scale Structural Similarity |
| IW-SSIM | 2011 | Information Content Weighted Structural Similarity Index |
| VIFp | 2004 | Visual Information Fidelity |
| FSIM | 2011 | Feature Similarity Index Measure |
| SR-SIM | 2012 | Spectral Residual Based Similarity |
| GMSD | 2013 | Gradient Magnitude Similarity Deviation |
| MS-GMSD | 2017 | Multi-Scale Gradient Magnitude Similarity Deviation |
| VSI | 2014 | Visual Saliency-induced Index |
| DSS | 2015 | DCT Subband Similarity Index |
| - | 2016 | Content Score |
| - | 2016 | Style Score |
| HaarPSI | 2016 | Haar Perceptual Similarity Index |
| MDSI | 2016 | Mean Deviation Similarity Index |
| LPIPS | 2018 | Learned Perceptual Image Patch Similarity |
| PieAPP | 2018 | Perceptual Image-Error Assessment through Pairwise Preference |
| DISTS | 2020 | Deep Image Structure and Texture Similarity |
| Acronym | Year | Metric |
|---|---|---|
| TV | 1937 | Total Variation |
| BRISQUE | 2012 | Blind/Referenceless Image Spatial Quality Evaluator |
| CLIP-IQA | 2022 | CLIP-IQA |
| Acronym | Year | Metric |
|---|---|---|
| IS | 2016 | Inception Score |
| FID | 2017 | Frechet Inception Distance |
| GS | 2018 | Geometry Score |
| KID | 2018 | Kernel Inception Distance |
| MSID | 2019 | Multi-Scale Intrinsic Distance |
| PR | 2019 | Improved Precision and Recall |
As part of our library we provide code to benchmark all metrics on a set of common Mean Opinon Scores databases. Currently we support several Full-Reference (TID2013, KADID10k and PIPAL) and No-Reference (KonIQ10k and LIVE-itW) datasets. You need to download them separately and provide path to images as an argument to the script.
Here is an example how to evaluate SSIM and MS-SSIM metrics on TID2013 dataset:
python3 tests/results_benchmark.py --dataset tid2013 --metrics SSIM MS-SSIM --path ~/datasets/tid2013 --batch_size 16Below we provide a comparison between Spearman's Rank Correlation Coefficient (SRCC) values obtained with PIQ and reported in surveys. Closer SRCC values indicate the higher degree of agreement between results of computations on given datasets. We do not report Kendall rank correlation coefficient (KRCC) as it is highly correlated with SRCC and provides limited additional information. We do not report Pearson linear correlation coefficient (PLCC) as it's highly dependent on fitting method and is biased towards simple examples.
For metrics that can take greyscale or colour images, c means chromatic version.
| TID2013 | KADID10k | PIPAL | |
|---|---|---|---|
| Source | PIQ / Reference | PIQ / Reference | PIQ / Reference |
| PSNR | 0.69 / 0.69 TID2013 | 0.68 / - | 0.41 / 0.41 PIPAL |
| SSIM | 0.72 / 0.64 TID2013 | 0.72 / 0.72 KADID10k | 0.50 / 0.53 PIPAL |
| MS-SSIM | 0.80 / 0.79 TID2013 | 0.80 / 0.80 KADID10k | 0.55 / 0.46 PIPAL |
| IW-SSIM | 0.78 / 0.78 Eval2019 | 0.85 / 0.85 KADID10k | 0.60 / - |
| VIFp | 0.61 / 0.61 TID2013 | 0.65 / 0.65 KADID10k | 0.50 / - |
| FSIM | 0.80 / 0.80 TID2013 | 0.83 / 0.83 KADID10k | 0.59 / 0.60 PIPAL |
| FSIMc | 0.85 / 0.85 TID2013 | 0.85 / 0.85 KADID10k | 0.59 / - |
| SR-SIM | 0.81 / 0.81 Eval2019 | 0.84 / 0.84 KADID10k | 0.57 / - |
| SR-SIMc | 0.87 / - | 0.87 / - | 0.57 / - |
| GMSD | 0.80 / 0.80 MS-GMSD | 0.85 / 0.85 KADID10k | 0.58 / - |
| VSI | 0.90 / 0.90 Eval2019 | 0.88 / 0.86 KADID10k | 0.54 / - |
| DSS | 0.79 / 0.79 Eval2019 | 0.86 / 0.86 KADID10k | 0.63 / - |
| Content | 0.71 / - | 0.72 / - | 0.45 / - |
| Style | 0.54 / - | 0.65 / - | 0.34 / - |
| HaarPSI | 0.87 / 0.87 HaarPSI | 0.89 / 0.89 KADID10k | 0.59 / - |
| MDSI | 0.89 / 0.89 MDSI | 0.89 / 0.89 KADID10k | 0.59 / - |
| MS-GMSD | 0.81 / 0.81 MS-GMSD | 0.85 / - | 0.59 / - |
| MS-GMSDc | 0.89 / 0.89 MS-GMSD | 0.87 / - | 0.59 / - |
| LPIPS-VGG | 0.67 / 0.67 DISTS | 0.72 / - | 0.57 / 0.58 PIPAL |
| PieAPP | 0.84 / 0.88 DISTS | 0.87 / - | 0.70 / 0.71 PIPAL |
| DISTS | 0.81 / 0.83 DISTS | 0.88 / - | 0.62 / 0.66 PIPAL |
| BRISQUE | 0.37 / 0.84 Eval2019 | 0.33 / 0.53 KADID10k | 0.21 / - |
| CLIP-IQA | 0.50 / - | 0.48 / - | 0.26 / - |
| IS | 0.26 / - | 0.25 / - | 0.09 / - |
| FID | 0.67 / - | 0.66 / - | 0.18 / - |
| KID | 0.42 / - | 0.66 / - | 0.12 / - |
| MSID | 0.21 / - | 0.32 / - | 0.01 / - |
| GS | 0.37 / - | 0.37 / - | 0.02 / - |
| KonIQ10k | LIVE-itW | |
|---|---|---|
| Source | PIQ / Reference | PIQ / Reference |
| BRISQUE | 0.22 / - | 0.31 / - |
| CLIP-IQA | 0.68 / 0.68 CLIP-IQA off | 0.64 / 0.64 CLIP-IQA off |
Unlike FR and NR IQMs, designed to compute an image-wise distance, the DB metrics compare distributions of sets of images.
To address these problems, we adopt a different way of computing the DB IQMs proposed in https://arxiv.org/abs/2203.07809.
Instead of extracting features from the whole images, we crop them into overlapping tiles of size 96 × 96 with stride = 32.
This pre-processing allows us to treat each pair of images as a pair of distributions of tiles, enabling further comparison.
The other stages of computing the DB IQMs are kept intact.
In PIQ we use assertions to raise meaningful messages when some component doesn't receive an input of the expected type.
This makes prototyping and debugging easier, but it might hurt the performance.
To disable all checks, use the Python -O flag: python -O your_script.py
See the open issues for a list of proposed features and known issues.
If you would like to help develop this library, you'll find more information in our contribution guide.
If you use PIQ in your project, please, cite it as follows.
@misc{kastryulin2022piq,
title = {PyTorch Image Quality: Metrics for Image Quality Assessment},
url = {https://arxiv.org/abs/2208.14818},
author = {Kastryulin, Sergey and Zakirov, Jamil and Prokopenko, Denis and Dylov, Dmitry V.},
doi = {10.48550/ARXIV.2208.14818},
publisher = {arXiv},
year = {2022}
}@misc{piq,
title={{PyTorch Image Quality}: Metrics and Measure for Image Quality Assessment},
url={https://github.com/photosynthesis-team/piq},
note={Open-source software available at https://github.com/photosynthesis-team/piq},
author={Sergey Kastryulin and Dzhamil Zakirov and Denis Prokopenko},
year={2019}
}Sergey Kastryulin - @snk4tr - snk4tr@gmail.com
Jamil Zakirov - @zakajd - djamilzak@gmail.com
Denis Prokopenko - @denproc - d.prokopenko@outlook.com