Official Code for the following paper:

X. Wang, A. Katsenou, and D. Bull. DIVA-VQA: Detecting Inter-Frame Variations in UGC Video Quality

Our paper was accepted by the IEEE International Conference on Image Processing. (ICIP 2025).

Try our online demo on Hugging Face 🤗: https://huggingface.co/spaces/xinyiW915/DIVA-VQA

We evaluate the performance of DIVA-VQA on five UGC datasets. Based on different pre-trained models and features, DIVA-VQA has two different versions of models:

• DIVA-VQA-B: Utilizes the SlowFast R50-3D and Swin-Base pre-trained models with a patch size of 16. The feature dimension is 9984.
• DIVA-VQA-L: Utilizes the SlowFast R50-3D and Swin-Large pre-trained models with a patch size of 16. The feature dimension is 11520.

More results and ablation studies can be found in correlation_result.ipynb.

Overview of the proposed framework with two core modules: Patch Difference Fragmentation and Spatio-Temporal Feature Extraction.

The comparison of running time (on GPU, averaged over ten runs) across different spatial resolutions:

The repository is built with Python 3.10 and can be installed via the following commands:

git clone https://github.com/xinyiW915/DIVA-VQA.git
cd DIVA-VQA
conda create -n divavqa python=3.10 -y
conda activate divavqa
pip install -r requirements.txt  

The corresponding UGC video datasets can be downloaded from the following sources:
LSVQ, KoNViD-1k, LIVE-VQC, YouTube-UGC, CVD2014.

The metadata for the experimented UGC dataset is available under ./metadata.

Once downloaded, place the datasets in any other storage location of your choice. Ensure that the videos_dir in the load_dataset function inside main_diva-vqa.py is updated accordingly.

Run the pre-trained models to evaluate the quality of a single video.

The model weights provided in ./log/reported_results/best_model for the fine-tuned version and ./log/best_model for the non-fine-tuned version, contain the best-performing saved weights from training. To evaluate the quality of a specific video, run the following command:

python test_demo.py 
    -device <DEVICE> 
    -is_finetune <True/False> 
    -save_path <MODEL_PATH> 
    -train_data_name <TRAIN_DATA_NAME> 
    -test_data_name <TEST_DATA_NAME>
    -test_video_path <DEMO_TEST_VIDEO> 
    -network_name <diva-vqa/diva-vqa_large>

Or simply try our demo video by running:

python test_demo.py 

Simply calculate the complexity of our demo video by running:

python test_demo_time_flops.py

Steps to train DIVA-VQA from scratch on different datasets.

Run the following command to extract features from videos:

python main_diva-vqa.py -database lsvq_train -feat_name diva-vqa_large -feature_save_path ../features/diva-vqa/

Train our model using extracted features:

python model_regression_simple.py -data_name youtube_ugc -network_name diva-vqa_large -feature_path ../features/ -save_path ../model/

For LSVQ, train the model using:

python model_regression.py -data_name lsvq_train -network_name diva-vqa_large -feature_path ../features/ -save_path ../model/

To fine-tune the pre-trained model on a new dataset, modify train_data_name to match the dataset used for training, and test_data_name to specify the dataset for fine-tuning. Also, modify network_name for different versions of DIVA-VQA and set the correct feature_path.

python model_finetune.py

This work was funded by the UKRI MyWorld Strength in Places Programme (SIPF00006/1) as part of my PhD study.

If you find this paper and the repo useful, please cite our paper 😊:

@article{wang2024relax,
      title={ReLaX-VQA: Residual Fragment and Layer Stack Extraction for Enhancing Video Quality Assessment},
      author={Wang, Xinyi and Katsenou, Angeliki and Bull, David},
      year={2024},
      eprint={2407.11496},
      archivePrefix={arXiv},
      primaryClass={eess.IV},
      url={https://arxiv.org/abs/2407.11496}, 
}
@article{wang2025diva,
      title={DIVA-VQA: Detecting Inter-Frame Variations in UGC Video Quality},
      author={Wang, Xinyi and Katsenou, Angeliki and Bull, David},
      booktitle={IEEE International Conference on Image Processing (ICIP 2025)}, 
      year={2025},
}

Xinyi WANG, [email protected]