GAMIVAL (GAMIng Video Quality EVALuator) is a no-reference video quality assessment model designed for cloud gaming videos. It extracts gaming-specific features in MATLAB and Python and trains a Support Vector Regressor (SVR) to predict perceptual quality. The method is introduced in IEEE Signal Processing Letters 2023.

• demo_compute_NSS_feats.m - MATLAB demo to compute Natural Scene Statistics (NSS) features.
• demo_compute_CNN_feats.py - Python script to extract 3D CNN features from video frames.
• combineFeature.m - merges NSS and CNN features into a single 2180-D descriptor.
• train_SVR.py / test_SVR.py - scripts to train or apply the SVR model.
• evaluate_bvqa_features_regression.py - evaluation utility used by run_all_bvqa_regression.sh.
• mos_files/ - metadata and MOS labels for supported datasets.
• feat_files/, models/, result/ - output folders for features, trained regressors, and results.

• MATLAB with access to ffmpeg for reading YUV videos.
• Python 3 with packages listed in requirements.txt (pip install -r requirements.txt).

1. Extract NSS Features

   demo_compute_NSS_feats.m

2. Extract CNN Features

   python demo_compute_CNN_feats.py --dataset_name LIVE-Meta-Gaming

Important: For LIVE-Meta-Gaming, all videos must be rescaled from TrueWidth × TrueHeight to DisplayWidth × DisplayHeight using bicubic interpolation prior to feature extraction. This matches the resolution used during subjective quality rating and ensures consistency in evaluation.

3. Combine Features

   combineFeature.m

4. Train or Evaluate

   bash run_all_bvqa_regression.sh        # batch evaluation
   # or
   python evaluate_bvqa_features_regression.py
   python train_SVR.py                     # train a custom model

5. Predict Quality Scores

   python test_SVR.py

GAMIVAL achieves state-of-the-art accuracy on the LIVE-Meta Mobile Cloud Gaming Database. Example SRCC/PLCC results are shown below.

Average feature extraction runtimes (1080p videos) are listed below.

A MATLAB and Python implementation of GAMIng Video Quality EVALuator (GAMIVAL), which is a new gaming-specific no reference video quality assessment model, proposed in IEEE SPL 2023. GAMIVAL achieves superior performance on the new LIVE-Meta Mobile Cloud Gaming Video Quality Assessment Database.

All videos, including training ones and testing ones, have their features (2180 features). The features are extracted first by a two-branch framework, which combines 1156 NSS features with 1024 CNN features. Then a support vector regressor is utilized to learn the feature-to-score mappings. The SVR parameters are optimized via a grid-search on the training set. Take LIVE-Meta-Mobile Cloud Gaming database for example, in the paper, 480 videos were used as training set, and other 120 videos were used as testing set. In application to Meta’s cloud game, we can use 600 videos as training set to gain a regressor as a quality predictor.

Schematic flow diagram of the GAMIVAL model. The top portion depicts the spatial and temporal NSS feature computations. The lower portion shows the CNN feature extraction process following NDNetGaming . All of the features are concatenated and utilized to train an SVR model.

demo_compute_NSS_feats.m

$ python demo_compute_CNN_feats.py --dataset_name LIVE-Meta-Gaming

Important: For LIVE-Meta-Gaming, all videos must be rescaled from TrueWidth × TrueHeight to DisplayWidth × DisplayHeight using bicubic interpolation prior to feature extraction. This matches the resolution used during subjective quality rating and ensures consistency in evaluation.

combineFeature.m

$ bash run_all_bvqa_regression.sh

or

$ python evaluate_bvqa_features_regression.py

$ python train_SVR.py

$ python test_SVR.py

Unit tests are located in the tests/ directory and can be executed with:

$ pytest

Box plots of PLCC, SRCC, and KRCC of evaluated BVQA algorithms on the LIVE-Meta MCG dataset over 1000 splits:

Speed was evaluated on the feature extraction function in all the algorithms. For GAMIVAL, speed was evaluated on demo_compute_NSS_feats.m and demo_compute_CNN_feats.py functions.

Scatter plots of SRCC of NR-VQA algorithms versus runtime on 1080p videos:

If you use this code for your research, please cite the following paper:

Y.-C. Chen, A. Saha, C. Davis, B. Qui, X. Wang, I. Katsavounidis, and A. C. Bovik, “Gamival : Video quality prediction on mobile cloud gaming content,” IEEE Signal Processing Letters, 2023, doi: 10.1109/LSP.2023.3255011.

@ARTICLE{10065464,
  author={Chen, Yu-Chih and Saha, Avinab and Davis, Chase and Qiu, Bo and Wang, Xiaoming and Gowda, Rahul and Katsavounidis, Ioannis and Bovik, Alan C.},
  journal={IEEE Signal Processing Letters},
  title={GAMIVAL: Video Quality Prediction on Mobile Cloud Gaming Content},
  year={2023},
  pages={1-5},
  doi={10.1109/LSP.2023.3255011}
}

• Yu-Chih Chen (berriechen@utexas.edu) – UT Austin
• Avinab Saha (avinab.saha@utexas.edu) – UT Austin
• Alan C. Bovik (bovik@ece.utexas.edu) – UT Austin