Main page — MSU Benchmark Collection

Home Participants Evaluation methodology How to participate Contact us

MSU Deinterlacer Benchmark — selecting the best deinterlacing filter

The most comprehensive comparison of deinterlacing methods

What’s new

26.11.2020 Beta-version Release
22.12.2020 Added Adobe Premiere Built-In, VS EEDI3, VS TDeintMod, MC Deinterlacer, EL MAB. Tuned Kernel Deinterlacer

Key features of the Benchmark

For Deinterlacing methods’ users
- Choose deinterlacing method that is the best for your speed and quality requirements
- Discover the newest deinterlacing methods’ achievements
For Researchers and Developers
- Quickly get comprehensive comparison results for your paper with our tables, visual comparison tools and performance plots
- Check the performance of your deinterlacing method on the complex cases

To submit deinterlacing method, please, follow 3 simple steps in the the Deinterlacer Submission section

We appreciate new ideas. Please, write us an e-mail to deinterlacer-benchmark@videoprocessing.ml

Leaderboard

The table below shows a comparison of deinterlacers by PSNR, SSIM metrics and by speed.

Click on the labels to sort the table

Rank	Name	PSNR	SSIM	FPS on CPU
1.0	MSU Deinterlacer	40.708	0.983	1.3
2.5	VapourSynth TDeintMod	39.916	0.977	50.29
3.0	NNEDI	39.625	0.978	1.91
4.0	Bob-Weave Deinterlacer	39.679	0.976	46.45
4.5	VapourSynth EEDI3	39.373	0.977	51.9
6.0	Real-Time Deep Deinterlacer	39.203	0.976	0.27
7.5	Bob	38.499	0.975	52.83
8.5	Weston 3-Field Deinterlacer	38.680	0.969	36.75
9.0	Kernel Deinterlacer (optimal parameters)	38.103	0.970	37.91
9.0	Elemental Live Low Latency Interpolation	38.056	0.972	Hardware Real-Time
11.0	YADIF	37.742	0.965	48.96
12.0	Elemental Live Motion Adaptive Interpolation	37.063	0.964	Hardware Real-Time
13.5	Kernel Deinterlacer	36.731	0.960	37.85
14.5	Studio Coast Pty vMix	36.990	0.950	Hardware Real-Time
14.5	Adobe Premiere Pro Built-In	36.092	0.958	43.82
16.0	Motion and Area Pixel Deinterlacer	35.415	0.950	2.15
16.5	Muksun Deinterlacer	35.444	0.949	1.95
18.0	PAL Interpolation	33.111	0.913	2.85
19.5	Elemental Live Motion Adaptive Blend	29.744	0.868	Hardware Real-Time
20.0	ASVZZZ Deinterlacer	27.499	0.902	1.9
20.5	Motion Compensation Deinterlacer	27.899	0.804	1.45

Full FrameRate Leaderboard

Rank	Name	PSNR	SSIM	FPS on CPU
1.0	MSU Deinterlacer	40.917	0.983	1.3
2.0	VapourSynth TDeintMod	40.071	0.978	50.29
3.5	VapourSynth EEDI3	39.547	0.978	51.9
4.0	Bob-Weave Deinterlacer	39.775	0.976	46.45
4.5	Real-Time Deep Deinterlacer	39.450	0.977	0.27
6.5	Bob	38.645	0.975	52.83
7.0	Weston 3-Field Deinterlacer	38.726	0.969	36.75
7.5	Elemental Live Low Latency Interpolation	37.908	0.972	Hardware Real-Time
9.0	YADIF	37.860	0.966	48.96
10.0	Elemental Live Motion Adaptive Interpolation	36.953	0.964	Hardware Real-Time
11.0	Studio Coast Pty vMix	32.942	0.931	Hardware Real-Time
12.0	Elemental Live Motion Adaptive Blend	29.747	0.868	Hardware Real-Time

Visualization

In this section you can see a frame, a crop from this frame, and also MSU VQMT PSNR Visualization of this crop.

Drag a red rectangle in the area which you want to crop, by default it is in the area with the worst PSNR

The area to compare on: Deinterlacer 1: Deinterlacer 2:

GT

In this row you can see VQMT PSNR Visualization

NNEDI

MSU Deinterlacer

VS TDeintMod

Charts

Highlight the plot region where you want to zoom in

Metric:

FPS is calculated on Intel-Core i7 10700K CPU

Metric:

The following plot shows difference between every method and Bob, because Bob is considered as the least complicated deinterlacing method

Metric:

Sequence №:

Sequence №: Metric:

Cross-PSNR

In this section you can see PSNR between the output of chosen deinterlacer and the others

Deinterlacer:

	PSNR
ASVZZZ	20.54547840460791
Bob	inf
Bob-Weave	42.07810541012702
Deep	41.164435818653764
EL LLI	43.87789597826898
EL MAI	38.01828007878827
Kernel	42.082198877630226
MAP	36.81130955719863
MSU	39.86926525444535
Muksun	37.93674757748442
NNEDI	43.222444723870076
PAL Interpolation	34.30022078877691
VMix	30.93623525121193
Weston 3-Field	42.526302640037
YADIF	39.61242962714242

Evaluation methodology

Dataset

Our dataset is constantly updated. Now we have 40 video sequences. Each sequence's length is 1 second. Resolution of all video sequences is 1920x1080. FPS varies from 24 to 60. TFF interlacing was used to get interlaced data from GT.

Click on this text to read how exactly our dataset was composed.

Initially we had about 30 videos from Vimeo 90k dataset, total length of which was about 1 hour. These videos included sports, panorama, news, landscapes, parts of movies, ads and other types of content.
We interlaced all these videos, and then measured PSNR between interlaced video and odd frames of gt.
From each video, we wanted to get 1 or 2 sequences, each 1 second long.
The sequences with the smallest mean PSNR were considered as the most “damaged”. Our hypothesis was that these sequences will be the hardest for deinterlacers.
Also we counted mean PSNR over all videos, and the sequences with the closest mean PSNR to mean-of-all PSNR was considered as the most “average”.
The sequences with the highest mean PSNR were considered as the most “undamaged”. These sequences were often a static shot with no moving objects.
We took 15 “damaged”, 20 “average”, and 5 “undamaged” sequences and put them together in one video, but separated by 10 black frames.
We don’t count PSNR on black frames. They were made for deinterlacers that use motion estimation (ME). Because of the black frames, ME-deinterlacer detects zero motion between two sequences and, therefore, doesn’t consider motion from the previous sequence while processing the current one. Also, we ignore the first 4 frames of each sequence while computing the overall mean. Again, we suppose that on the first 4 frames ME-deinterlacers are collecting information about motion and don't show their best.

Metrics

We compare RGB frames via 2 metrics - PSNR and SSIM. PSNR and SSIM are measured over the Y-component.

For each video sequence, we take the average PSNR and SSIM over all frames. We decided to choose these metrics because they proved themselves to be among the best metrics to show quality loss.

We decided to measure these metrics over Y component because YUV is the most popular type of colorspace nowadays, but there are still a lot of versions of YUV (e.g. yuv444p, yuv420p, yuv420p12le). In these versions U and V components are different, that’s why we measure only Y component. Also, there are a lot of other color spaces, which use Y component (e.g. YCbCr, YPbPr, UYVY, ...). Finally, we can easily compute Y component from other color spaces, such as RGB or gray.

Here is the plot of the PSNR difference between PSNR-Y and PSNR-RGB. As we can see here, the difference is rather negligible.

Validation of deinterlacers' outputs

Another important direction of our work is to control the outputs of deinterlacers. Sometimes, it can convert colorspace, work in BFF mode instead of TFF, or maybe have bugs.

Click on this text to read how exactly we validate deinterlacers' outputs.

The main criteria is that the PSNR between GT fields and the same fields in deinterlaced video must be equal to infinity. To control this, we make Top-Fields and Bottom-Fields plot on every second frame of GT video and deinterlaced video.

Here is the sample plot for Bob-Weave Deinterlacer. This deinterlacer passed the validation.

PSNR between Bottom-Fields of every deinterlaced frame and GT frame are equal to infinity, so we substitute infinity by zero. This means, that bottom field exactly matches the corresponding field in GT sequence.

Another sample plot, for MSU Deinterlacer, which had problems with colorspace.

As we can see here, Bottom-Field PSNR is about 100, but not infinity. In such cases, we make the following steps:

Convert GT-data to the colorspace of the deinterlacer.
If we don’t know (can’t guess) which colorspace the deinterlacer uses, we form the LookUp-Table from the fields of the deinterlacer output and GT data, that must be equal to each. Then, we use this LookUp-Table to map GT to the colorspace of the deinterlacer.
In some cases, it is impossible to precisely determine LookUp-Table because the mapping is neither injective, nor surjective. In such hard cases, we just choose a GT sample with the highest PSNR from the previous 2 steps.

Also, we provide MSU Video Quality Measurement Tool (VQMT) PSNR visualization of the deinterlacer output. As the last step of the validation process, we check, that the VQMT PSNR visualization is striped. It should be so, because even/odds rows must exactly match the corresponding GT rows.

Here in the example of correct VQMT PSNR visualization output:

And, finally, let us make a closer look at it:

As we can see, the output is striped, so this means that the deinterlacer passed the last validation step.

Deinterlacer Submission

There are 3 easy steps to submit:

Download the interlaced video here.
We have more available formats, if YUV is not suitable. Just click on this text

There are 5 available options:

Deinterlace downloaded video
The details, which may help you
You can also send us any executable file or code of your deinterlacer, in order not to deinterlace video by yourself

Send us an email to deinterlacer-benchmark@videoprocessing.ml with the following information:

Contacts

For questions and propositions, please contact us: deinterlacer-benchmark@videoprocessing.ml

05 Nov 2020