Awesome
SimpleVQA
A Deep Learning based No-reference Quality Assessment Model for UGC Videos
Description
This is a repository for the model proposed in the paper "A Deep Learning based No-reference Quality Assessment Model for UGC Videos". Arxiv Version ACM MM 2022 Version
Usage
Install Requirements
pytorch
opencv
scipy
pandas
torchvision
torchvideo
Download databases
Train models
- Extract video frames
python -u extract_frame_LSVQ.py >> logs/extract_frame_LSVQ.log
- Extract motion features
CUDA_VISIBLE_DEVICES=0 python -u extract_SlowFast_features_LSVQ.py \
--database LSVQ \
--model_name SlowFast \
--resize 224 \
--feature_save_folder LSVQ_SlowFast_feature/ \
>> logs/extracted_LSVQ_SlowFast_features.log
- Train the model
CUDA_VISIBLE_DEVICES=0 python -u train_baseline.py \
--database LSVQ \
--model_name UGC_BVQA_model \
--conv_base_lr 0.00001 \
--epochs 10 \
--train_batch_size 8 \
--print_samples 1000 \
--num_workers 6 \
--ckpt_path ckpts \
--decay_ratio 0.9 \
--decay_interval 2 \
--exp_version 0 \
--loss_type L1RankLoss \
--resize 520 \
--crop_size 448 \
>> logs/train_UGC_BVQA_model_L1RankLoss_resize_520_crop_size_448_exp_version_0.log
Test the model
You can download the trained model via Google Drive.
Test on the public VQA database
CUDA_VISIBLE_DEVICES=0 python -u test_on_pretrained_model.py \
--database KoNViD-1k \
--train_database LSVQ \
--model_name UGC_BVQA_model \
--feature_type SlowFast \
--trained_model ckpts/UGC_BVQA_model.pth \
--num_workers 6 \
>> logs/test_on_KoNViD-1k_train_on_LSVQ.log
Test on a single video
CUDA_VISIBLE_DEVICES=0 python -u test_demo.py \
--method_name single-scale \
--dist videos/2999049224_original_centercrop_960x540_8s.mp4 \
--output result.txt \
--is_gpu \
>> logs/test_demo.log
Citation
If you find this code is useful for your research, please cite:
@inproceedings{sun2022a,
title = {A Deep Learning Based No-Reference Quality Assessment Model for UGC Videos},
author = {Sun, Wei and Min, Xiongkuo and Lu, Wei and Zhai, Guangtao},
booktitle={Proceedings of the 30th ACM International Conference on Multimedia},
year = {2022},
pages = {856–865},
}