Skip to content

A resource list and performance benchmark for blind video quality assessment (BVQA) models on user-generated content (UGC) datasets. [IEEE TIP'2021] "UGC-VQA: Benchmarking Blind Video Quality Assessment for User Generated Content", Zhengzhong Tu, Yilin Wang, Neil Birkbeck, Balu Adsumilli, Alan C. Bovik

License

Notifications You must be signed in to change notification settings

vztu/BVQA_Benchmark

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

32 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

BVQA_Benchmark

This is a resource list for blind video quality assessment (BVQA) models on user-generated databases, i.e., the UGC-VQA problem studied in our paper [IEEE TIP2021] UGC-VQA: Benchmarking blind video quality assessment for user generated content. IEEEXplore

The following content include datasets, models & codes, performance benchmark and leaderboard.

Maintained by: Zhengzhong Tu

👍 Any suggestion or idea is welcomed. Please see Contributing

  • Updates [02-31-2021] Added VQA mainpage at papers-with-code here: video-quality-assessment
  • Updates [10-21-2021] All the features I used in the paper can be downloaded here: Google Drive

Contents

Evaluate Your Own Model

Extract features in the form of NxM matrix (N:#samples, M:#features) on a given VQA dataset and save it in data/ directory. Let metadata file be stored in the same folder with MOSs in the same order as your feature matrix (We have already provided the MOS arrays of three UGC datasets). The evaluate_bvqa_features.py evaluates the extracted features via 100 random train-test splits and reports the median (std) SRCC/KRCC/PLCC/RMSE performances. Note that it is not applicable to deep learning models (feature-based model only).

Pre-requisites

  • python3
  • sklearn

Demo evaluations (BRISUQE on KoNViD-1k)

$ python3 src/evaluate_bvqa_features.py

Custom usage with your own model on given dataset

$ python3 src/evaluate_bvqa_features.py [-h] [--model_name MODEL_NAME]
                                   [--dataset_name DATASET_NAME]
                                   [--feature_file FEATURE_FILE]
                                   [--mos_file MOS_FILE] [--out_file OUT_FILE]
                                   [--color_only] [--log_short] [--use_parallel]
                                   [--num_iterations NUM_ITERATIONS]
                                   [--max_thread_count MAX_THREAD_COUNT]

UGC-VQA Datasets

BVQA Dataset Download Paper
KoNViD-1k (2017) KoNViD-1k Hosu et al. QoMEX'17
LIVE-VQC (2018) LIVE-VQC Sinno et al. TIP'19
YouTube-UGC (2019) YouTube-UGC Wang et al. MMSP'19
LIVE-FB-LSVQ (2021) LIVE-FB-LSVQ Ying et al. CVPR'21

BIQA / BVQA Models

BIQA

Model Download Paper
BRISQUE BRISQUE Mittal et al. TIP'12
NIQE NIQE Mittal et al. TIP'13
ILNIQE ILNIQE Zhang et al. TIP'15
GM-LOG GM-LOG Xue et al. TIP'14
HIGRADE HIGRADE Kundu et al. TIP'17
FRIQUEE FRIQUEE Ghadiyaram et al. JoV'17
CORNIA BIQA_Toolbox Ye et al. CVPR'12
HOSA BIQA_Toolbox Xu et al. TIP'16
KonCept 512 koniq, koniq-PyTorch Hosu et al. TIP'20
PaQ-2-PiQ PaQ-2-PiQ, paq2piq-PyTorch Ying et al. CVPR'20

BVQA

Model Download Paper
VIIDEO VIIDEO Mittal et al. TIP'16
V-BLIINDS V-BLIINDS Saad et al. TIP'14
TLVQM nr-vqa-consumervideo Korhenen et al. TIP'19
VSFA VSFA Li et al. MM'19
NSTSS NRVQA-NSTSS Dendi et al. TIP'20
VIDEVAL VIDEVAL Tu et al. TIP'21
MDTVSFA MDTVSFA Li et al. IJCV'21
RAPIQUE RAPIQUE Tu et al. OJSP'21
PatchVQ PatchVQ Ying et al. CVPR'21
CoINVQ CoINVQ Wang et al. CVPR'21

Performance Benchmark

Regression Results

Median SRCC (std SRCC) of 100 random train-test (80%-20%) splits.

Methods KoNViD-1k LIVE-VQC YouTube-UGC All-Combined
BRISQUE 0.6567 (0.0351) 0.5925 (0.0681) 0.3820 (0.0519) 0.5695 (0.0289)
NIQE 0.5417 (0.0347) 0.5957 (0.0571) 0.2379 (0.0487) 0.4622 (0.0313)
IL-NIQE 0.5264 (0.0294) 0.5037 (0.0712) 0.2918 (0.0502) 0.4592 (0.0307)
GM-LOG 0.6578 (0.0324) 0.5881 (0.0683) 0.3678 (0.0589) 0.5650 (0.0295)
HIGRADE 0.7206 (0.0302) 0.6103 (0.0680) 0.7376 (0.0338) 0.7398 (0.0189)
FRIQUEE 0.7472 (0.0263) 0.6579 (0.0536) 0.7652 (0.0301) 0.7568 (0.0237)
CORNIA 0.7169 (0.0245) 0.6719 (0.0473) 0.5972 (0.0413) 0.6764 (0.0216)
HOSA 0.7654 (0.0224) 0.6873 (0.0462) 0.6025 (0.0344) 0.6957 (0.0180)
VGG-19 0.7741 (0.0288) 0.6568 (0.0536) 0.7025 (0.0281) 0.7321 (0.0180)
ResNet-50 0.8018 (0.0255) 0.6636 (0.0511) 0.7183 (0.0281) 0.7557 (0.0177)
KonCept512 0.7349 (0.0252) 0.6645 (0.0523) 0.5872 (0.0396) 0.6608 (0.0221)
PaQ-2-PiQ 0.6130 (0.0325) 0.6436 (0.0457) 0.2658 (0.0473) 0.4727 (0.0298)
VIIDEO 0.2988 (0.0561) 0.0332 (0.0856) 0.0580 (0.0536) 0.1039 (0.0349)
V-BLIINDS 0.7101 (0.0314) 0.6939 (0.0502) 0.5590 (0.0496) 0.6545 (0.0232)
TLVQM 0.7729 (0.0242) 0.7988 (0.0365) 0.6693 (0.0306) 0.7271 (0.0189)
VIDEVAL 0.7832 (0.0216) 0.7522 (0.0390) 0.7787 (0.0254) 0.7960 (0.0151)
VSFA 0.755 (0.025) - - -
NSTSS 0.6417 - - -
VIDEVAL+KonCept512 0.8149 (0.0194) 0.7849 (0.0440) 0.8083 (0.0232) 0.8123 (0.0163)
MDTVSFA 0.7812 (0.0278) 0.7382 (0.0357) - -
RAPIQUE 0.8031 0.7548 0.7591 0.8070
PatchVQ 0.791 0.827 - -
CoINVQ 0.767 - 0.816 -

The median PLCC (std PLCC) of 100 random train-test (80%-20%) splits.

Model KoNViD-1k LIVE-VQC YouTube-UGC All-Combined
BRISQUE 0.6576 (0.0342) 0.6380 (0.0632) 0.3952 (0.0486) 0.5861 (0.0272)
NIQE 0.5530 (0.0337) 0.6286 (0.0512) 0.2776 (0.0431) 0.4773 (0.0287)
IL-NIQE 0.5400 (0.0337) 0.5437 (0.0707) 0.3302 (0.0579) 0.4741 (0.0280)
GM-LOG 0.6636 (0.0315) 0.6212 (0.0636) 0.3920 (0.0549) 0.5942 (0.0306)
HIGRADE 0.7269 (0.0287) 0.6332 (0.0652) 0.7216 (0.0334) 0.7368 (0.0190)
FRIQUEE 0.7482 (0.0257) 0.7000 (0.0587) 0.7571 (0.0324) 0.7550 (0.0226)
CORNIA 0.7135 (0.0236) 0.7183 (0.0420) 0.6057 (0.0399) 0.6974 (0.0202)
HOSA 0.7664 (0.0207) 0.7414 (0.0410) 0.6047 (0.0347) 0.7082 (0.0167)
VGG-19 0.7845 (0.0246) 0.7160 (0.0481) 0.6997 (0.0281) 0.7482 (0.0176)
ResNet-50 0.8104 (0.0229) 0.7205 (0.0434) 0.7097 (0.0276) 0.7747 (0.0167)
KonCept512 0.7489 (0.0240) 0.7278 (0.0464) 0.5940 (0.0412) 0.6763 (0.0227)
PaQ-2-PiQ 0.6014 (0.0338) 0.6683 (0.0445) 0.2935 (0.0490) 0.4828 (0.0293)
VIIDEO 0.3002 (0.0539) 0.2146 (0.0903) 0.1534 (0.0498) 0.1621 (0.0355)
V-BLIINDS 0.7037 (0.0301) 0.7178 (0.0500) 0.5551 (0.0465) 0.6599 (0.0234)
TLVQM 0.7688 (0.0238) 0.8025 (0.0360) 0.6590 (0.0302) 0.7342 (0.0180)
VIDEVAL 0.7803 (0.0223) 0.7514 (0.0420) 0.7733 (0.0257) 0.7939 (0.0157)}
VSFA 0.744 (0.029) - - -
NSTSS 0.6531 - - -
VIDEVAL+KonCept512 0.8169 (0.0179) 0.8010 (0.0398) 0.8028 (0.0234) 0.8168 (0.0128)
MDTVSFA 0.7856 (0.0240) 0.7728 (0.0351) - -
RAPIQUE 0.8175 0.7863 0.7684 0.8229
PatchVQ 0.786 0.837 - -
CoINVQ 0.764 - 0.802 -

Contributing

Please feel free to send an issue or pull requests or email me to add links or new results.

Citation

Should you find this repo useful to your research, we sincerely appreciate it if you cite our papers:blush::

@article{tu2020ugc,
  title={UGC-VQA: Benchmarking Blind Video Quality Assessment for User Generated Content},
  author={Tu, Zhengzhong and Wang, Yilin and Birkbeck, Neil and Adsumilli, Balu and Bovik, Alan C},
  journal={arXiv preprint arXiv:2005.14354},
  year={2020}
}

@inproceedings{tu2020comparative,
  title={A Comparative Evaluation Of Temporal Pooling Methods For Blind Video Quality Assessment}, 
  author={Z. {Tu} and C. -J. {Chen} and L. -H. {Chen} and N. {Birkbeck} and B. {Adsumilli} and A. C. {Bovik}},
  booktitle={2020 IEEE International Conference on Image Processing (ICIP)},  
  year={2020},
  pages={141-145},
  doi={10.1109/ICIP40778.2020.9191169}
}

About

A resource list and performance benchmark for blind video quality assessment (BVQA) models on user-generated content (UGC) datasets. [IEEE TIP'2021] "UGC-VQA: Benchmarking Blind Video Quality Assessment for User Generated Content", Zhengzhong Tu, Yilin Wang, Neil Birkbeck, Balu Adsumilli, Alan C. Bovik

Topics

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages