This repository contains the official implementation of the Cut-FUNQUE model proposed in the following paper.

1. A. K. Venkataramanan, C. Stejerean, I. Katsavounidis, H. Tmar and A. C. Bovik, "Cut-FUNQUE: An Objective Quality Model for Compressed Tone-Mapped High Dynamic Range Videos," arXiv preprint 2024.

Cut-FUNQUE achieves state-of-the-art quality prediction accuracy on tone-mapped and compressed high dynamic range videos, at a fraction of the computational complexity of SOTA models like MSML.

1. A novel perceptually uniform encoding of color signals (PUColor) that we use to represent both HDR and SDR color stimuli in a common domain. In this manner, PUColor enables the meaningful comparison of stimuli across dynamic ranges, which is essential when comparing HDR and SDR videos.
2. A binned-weighting approach to separately handle image regions having different visual characteristics such as brightness, contrast, and temporal complexity.
3. Novel statistical similarity measures of visual quality to overcome the limitations of pixel-wise comparisons across dynamic ranges.

The Cut-FUNQUE model achieves SOTA accuracy at a fraction of the existing SOTA MSML!

Create and activate a virtual environment using

python3 -m virtualenv .venv
source .venv/bin/activate

Install all required dependencies

python3 -m pip install -r requirements.txt

To compute Cut-FUNQUE features from one video pair, use the command

python3 extract_features.py --ref_video <path to reference video> --dis_video <path to distorted video>

For more options, run

python3 extract_features.py --help

First, define a subjective dataset file using the same format as that in datasets/. The dataset file provided here is that of the LIVE-TMHDR dataset, which was used to benchmark Cut-FUNQUE.

Then, run

python3 extract_features_from_dataset.py --dataset <path to dataset file> --processes <number of parallel processes to use>

Note: This command computes features and saves the results to disk. It does not print any features. Saved features may be used for downstream tasks - example below

To evaluate features using content-separated random cross-validation, run

python3 crossval_features_on_dataset.py --dataset <path to dataset file> --splits <number of random train-test splits> --processes <number of parallel processes to use>

Note: This command may be run without running extract_features_from_dataset.py first. In that case, features will be extracted and saved first, before performing cross-validation

This script is an example of down-stream tasks that can be performed easily after feature extraction.

If you encounter any bugs, or you have suggestions or contributions to improve this code base, please raise an Issue!