Skip to content

Latest commit

 

History

History
executable file
·
71 lines (45 loc) · 3.08 KB

README.md

File metadata and controls

executable file
·
71 lines (45 loc) · 3.08 KB

HIDRO-VQA: High Dynamic Range Oracle for Video Quality Assessment

This is the official repository of the paper HIDRO-VQA

arXiv

HIDRO-VQA: High Dynamic Range Oracle for Video Quality Assessment
Shreshth Saini*, Avinab Saha*, and Alan C. Bovik
3rd Workshop on Image/Video/Audio Quality in Computer Vision and Generative AI
WACV 2024

Usage

The code has been tested on Linux systems with Python 3.9. Please refer to requirements.txt for installing dependent packages.

Running HIDRO-VQA

In order to obtain quality-aware features or to start HDR quality-aware fine-tuning, checkpoints need to be downloaded. Download the checkpoint folder from the link and save them to the checkpoints folder.

Obtaining HIDRO-VQA Features on the LIVE HDR Database

For obtaining HIDRO-VQA features, the following command can be used. The features are saved in '.npy' format. It assumes the videos are stored in raw YUV 10-bit format, upscaled to 4K. Please change the path of the videos in line 97.

python demo_hidro_vqa_feats.py

Training HIDRO-VQA

Data Processing

Please take a look at the DATA folder for steps on how to prepare data for HIDRO-VQA pre-training.

Training Model

Training with multiple GPUs using Distributed training

Run the following commands on different terminals concurrently. Please update the folder location of HDR Frames in the data_loader file located in modules/data_loader.py (Lines 119,137).

CUDA_VISIBLE_DEVICES=0 python train.py --nodes 3 --nr 0 --batch_size 256 --lr 0.1 --epochs 25
CUDA_VISIBLE_DEVICES=1 python train.py --nodes 3 --nr 1 --batch_size 256 --lr 0.1 --epochs 25
CUDA_VISIBLE_DEVICES=2 python train.py --nodes 3 --nr 2 --batch_size 256 --lr 0.1 --epochs 25

Note that in distributed training, batch_size value will be the number of images to be loaded on each GPU.

Training Regressor

After HIDRO-VQA model pre-training, an SVR is trained using HIDRO_VQA features and corresponding ground truth quality scores from LIVE-HDR using the following command. It assumes features from each video is extracted (per frame) and stored using individual numpy files in a folder.

python train_svr.py --score_file <score_csv_file> --feature_folder <feature_folder_path> --train_and_test

Contributing

Contributions are welcome! If you find any issues or have suggestions for improvements, please open an issue or submit a pull request.

Citation

@InProceedings{Saini_2024_WACV,
author    = {Saini, Shreshth and Saha, Avinab and Bovik, Alan C.},
title     = {HIDRO-VQA: High Dynamic Range Oracle for Video Quality Assessment},
booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops},
month     = {January},
year      = {2024},
pages     = {469-479}}

License

This project is licensed under the MIT License.