MSU Video Super Resolution Benchmark 2021 — find the best upscaler

Discover the newest methods and find the most appropriate method for your tasks

Video group head: Dr. Dmitriy Vatolin

Measurements, analysis:

Anastasia Kirillova,
Eugene Lyapustin

What’s new

26.08.2021 Added RBPN, iSeeBetter, TMNet.
11.08.2021 Added ESRGAN, Real-ESRGAN, Real-ESRNet, RSDN, TGA.
15.07.2021 Improved Visualization section to be more user-friendly and added plots with the metric difference on BI and BD degradation to Leaderboard section.
21.05.2021 Added new plots to Charts section.
26.04.2021 Beta-version Release

Key features of the Benchmark

New metrics for the detail restoration quality
- Check methods’ ability to restore real details
The most complex content for restoration task: faces, text, QR-codes, car numbers, unpatterned textures, small details
- See plots and visualizations for particular content types
Different degradation types to lower the resolution: bicubic interpolation (BI) and Gaussian blurring and downsampling (BD), with and without noise
- Many methods use only one degradation type for their training datasets (e.g. bicubic interpolation) and do not work well on others. Choose the method that didn’t overfit the test dataset
Subjective comparison of new and popular Super-Resolution methods

The test-stand of the benchmark: the captured frame and the frame divided by content types.
See methodology for more information.

Leaderboard

The table below shows a comparison of Video Super Resolution methods by subjective score and a few objective metrics on Bicubic interpolation (BI). Default ranking is by subjective score.
You can click on the model's name in the table to read information about the method. You can see information about all participants here.

Metrics: Subjective assessment — read about,
ERQAv1.0 — read about,
PSNR-Y — read about,
SSIM-Y — read about,
QRCRv1.0 — read about,
CRRMv1.0 — read about,
FPS — read about.

Click on the labels to sort the table

Rank	Model	Subjective	ERQAv1.0	QRCRv1.0	SSIM-Y**	CRRMv1.0	PSNR-Y**	FPS(s)
16	D3Dnet^[16]	3.283	0.674	0.549	0.876	0.975	29.703	24.238¹
2	DBVSR^[2]	5.247	0.737	0.629	0.894	0.992	31.071	TBP*
15	DUF-16L^[15]	3.393	0.641	0.549	0.828	0.968	24.606	1.653¹
13	DUF-28L^[13]	3.571	0.645	0.549	0.830	0.993	25.852	2.392¹
5	DynaVSR-R^[5]	4.403	0.709	0.557	0.865	0.997	28.377	5.664¹
20	DynaVSR-V^[20]	2.554	0.643	0.549	0.864	0.997	29.011	6.660¹
22	ESPCN^[22]	0.214	0.521	0.000	0.811	0.948	26.714	TBP*
10	ESRGAN^[10]	3.674	0.735	0.000	0.808	0.998	27.330	0.996¹
4	LGFN^[4]	4.715	0.740	0.629	0.898	0.996	31.291	1.499¹
1	RBPN^[1]	5.325	0.746	0.629	0.899	0.998	31.407	TBP*
12	RRN-10L^[12]	3.587	0.627	0.557	0.790	0.989	24.252	0.390¹
17	RRN-5L^[17]	3.245	0.617	0.549	0.789	0.989	23.786	0.365¹
7	RSDN^[7]	3.845	0.667	0.619	0.826	0.985	25.321	TBP*
11	Real-ESRGAN^[11]	3.652	0.663	0.000	0.774	0.942	24.441	0.991¹
21	Real-ESRnet^[21]	1.830	0.598	0.000	0.824	0.951	27.195	0.982¹
14	RealSR^[14]	3.520	0.690	0.000	0.767	0.886	25.989	TBP*
18	SOF-VSR-BD^[18]	3.110	0.647	0.557	0.831	0.971	25.986	1.430¹
19	SOF-VSR-BI^[19]	3.041	0.660	0.557	0.872	0.974	29.381	1.750¹
9	TDAN^[9]	3.718	0.706	0.609	0.883	0.994	30.244	TBP*
8	TGA^[8]	3.832	0.669	0.549	0.831	0.987	25.786	1.417¹
6	TMNet^[6]	4.298	0.712	0.549	0.885	0.997	30.364	TBP*
3	iSeeBetter^[3]	5.122	0.748	0.629	0.896	1.000	31.104	22.275¹

*To Be Published
**Shifted version (see methodology for more information)
¹ measured by MSU
² measured by participants

The absolute difference between metric values on BI and BD degradation models

Metric:

Click here to see table with metrics on data prepared with BD degradation

Click on the labels to sort the table

Rank	Model	Subjective	ERQAv1.0	QRCRv1.0	SSIM-Y**	CRRMv1.0	PSNR-Y**	FPS(s)
16	D3Dnet^[16]	3.283	0.538	0.000	0.831	0.956	27.093	24.238¹
2	DBVSR^[2]	5.247	0.589	0.557	0.843	0.964	27.664	TBP*
15	DUF-16L^[15]	3.393	0.533	0.000	0.820	0.995	24.969	1.653¹
13	DUF-28L^[13]	3.571	0.534	0.000	0.823	0.976	26.365	2.392¹
5	DynaVSR-R^[5]	4.403	0.690	0.619	0.882	0.977	30.182	5.664¹
20	DynaVSR-V^[20]	2.554	0.666	0.629	0.875	0.986	29.724	6.660¹
22	ESPCN^[22]	0.214	0.411	0.000	0.784	0.940	25.672	TBP*
10	ESRGAN^[10]	3.674	0.506	0.000	0.797	0.964	26.115	0.996¹
4	LGFN^[4]	4.715	0.618	0.629	0.853	0.966	28.020	1.499¹
1	RBPN^[1]	5.325	0.595	0.629	0.847	0.966	27.765	TBP*
12	RRN-10L^[12]	3.587	0.517	0.609	0.798	0.977	25.001	0.390¹
17	RRN-5L^[17]	3.245	0.505	0.549	0.793	0.977	24.382	0.365¹
7	RSDN^[7]	3.845	0.554	0.557	0.826	0.978	25.767	TBP*
11	Real-ESRGAN^[11]	3.652	0.656	0.000	0.764	0.950	24.239	0.991¹
21	Real-ESRnet^[21]	1.830	0.591	0.000	0.822	0.955	27.194	0.982¹
14	RealSR^[14]	3.520	0.588	0.000	0.751	0.948	25.406	TBP*
18	SOF-VSR-BD^[18]	3.110	0.536	0.549	0.826	0.954	26.808	1.430¹
19	SOF-VSR-BI^[19]	3.041	0.527	0.000	0.826	0.955	27.004	1.750¹
9	TDAN^[9]	3.718	0.550	0.000	0.831	0.964	27.057	TBP*
8	TGA^[8]	3.832	0.554	0.557	0.828	0.976	26.247	1.417¹
6	TMNet^[6]	4.298	0.537	0.000	0.827	0.964	26.918	TBP*
3	iSeeBetter^[3]	5.122	0.596	0.557	0.848	0.969	27.712	22.275¹

*To be published
**Shifted version (see methodology for more information)
¹ measured by MSU
² measured by participants

Correlation of metrics with subjective assessment

Correlation method:

Charts

In this section, you can see barcharts and speed-to-performance plots. You can choose metric, motion, content, and degradation type, see tests with or without noise.
You can see information about all participants here.

Metric: Test: Content:

Highlight the plot region where you want to zoom in

Visualization

In this section, you can choose a part of the frame with particular content, see a cropped piece from this, MSU VQMT PSNR* Visualization, and ERQAv1.0 Visualization for this crop. In part "QR-codes" codes, which can be detected, are surrounded by a blue rectangle. You can see information about all participants here.
*We visualize shifted PSNR metric by applying MSU VQMT PSNR Visualization to frames with optimal shift for PSNR.

Frame: Content:

Model 1: Model 2: Model 3:

Drag a red rectangle in the area, which you want to crop.

GT

D3Dnet

DBVSR

DUF-16L

Your method submission

Verify the restoration ability of your VSR algorithm and compare it with state-of-the-art solutions.
You can see information about all other participants here.

1. Download input data		Download input low-resolution videos as sequences of frames in .png format There are 2 available options: a. Download 1 folder with all videos, joined in one sequence here. Neighboring videos are separated by 10 black frames, which will be skipped for evaluation. b. If you worry this strategy can decrease your performance, you can download 12 folders with 100 frames each here.

2. Apply your algorithm		Restore high-resolution frames with your algorithm. You can also send us the code of your method or the executable file and we will run it ourselves.

3. Send us result		Send us an email to vsr-benchmark@videoprocessing.ai with the following information: A. Name of your method that will be specified in our benchmark B. Link to the cloud drive (Google Drive, OneDrive, Dropbox, etc.), containing output frames. Check that the number of output frames matches the number of input frames: 1310 frames in one folder for download option (a) 12 folders with 100 frames each for download option (b) C. (Optional) Execution time of your algorithm and information about used GPU D. (Optional) Any additional information about the method: 1. Full name of your model 2. The parameter set that was used 3. Any other additional information 4. A link to the code of your model, if it is available 5. A link to the paper about your model 6. Any characteristics of your model's architecture (e.g. motion compensation, propagation, fusion)

Contacts

For questions and propositions, please contact us: vsr-benchmark@videoprocessing.ai

26 Apr 2021