Deinterlacing Challenge 2019

Musicvid wrote on 1/13/2019, 3:34 PM

Note: If you downloaded the MP4 test file below before 1/15/19, please do so again. The source half of the image (on the left) is improved.

 

Nick Hope wrote on 10/3/2016, 11:50 PM

How about a new "challenge" where we compare:

Vegas interpolate

Vegas smart adaptive

Handbrake decomb

QTGMC

Vegas Yadif plugin

Anything else

Here's one:

Download this test (mp4) (Revised 1/14/2019)

https://drive.google.com/file/d/1zB_oEF0KCwV0RT3kfhbv-grIiaBzrKlK/view?usp=sharing

Download AVCHD 1080 59.94i CG Source (roll your own tests)

https://drive.google.com/file/d/1Le5Ocyb0fJpMHucuxyDzKNDwSOZsOGl_/view?usp=sharing

Background:

https://www.vegascreativesoftware.info/us/forum/vegas-to-the-web-for-the-videophile-a-tutorial--83144/

Considerations:

Subjective visual comparisons of footage-in-the-wild using various encoder deinterlacers are fraught with variables -- perhaps the two biggest being the observer and the viewing screen. In addition, it has proven difficult to find real-world footage containing sustained sharp oblique or radial motion, which are the truest test of decombing as compared to deinterlacing, the latter being at its best with lateral subject and camera movement. Stringer's enduring "Driving Along" clip is a good example of this.

For that reason, I set out to contrive a synthetic deinterlacing benchmark, and it has turned out to be a lot of fun. The idea of a whirling Lissajous was borne of Glenmorangie, and not anything more sinister.

Disclaimer:

The content image is generated by Lissajous3D, a shareware program. No corporate logos were harmed in the making of this movie. My benchmarking tools (signature) are free for noncommercial use, as always..

http://robertinventor.com/software/main/acceptable_use.htm

Usage and Viewing Environment:

-- Comparisons are valid only at 1920x1080 30p resolution at 1:1 Scaling.

-- Edge Inspection: Freeze or Grab a frame, magnify, and compare with other solutions. You will see some marked differences.

Download PNG Here (1200x1400):

https://drive.google.com/file/d/1Q_5Uj1vSX0JAmu21kzAzMV-cFIo5QcvL/view?usp=sharing

-- Temporal (Time / Motion) Testing -- This is where you unpause, let it loop, stand back for a while, and form your own impressions. You will find straightaway that every playback environment will look different with every deinterlacer, so your decisions with regard to smoothness vs. sharpness will mostly rely on the screen environment you most often watch..

-- Next, of course, come the subjective decisions based on the way you want to see your own video. It's also good to ask the opinions of others.

 

Comments

john_dennis wrote on 1/13/2019, 6:45 PM

On a somewhat parallel track, I've been testing wwaag's Tool for Measuring Render Quality. After spending time on the usual render suspects, I decided to try an interlaced file. I chose the Mainconcept Blu-ray 1920x1080-60i, 25 Mbps video stream since I've used that template in the past. I started with a 3840x2160-29.97p file and rendered at default template settings except I had to lower the maximum bit rate to 39 Mbps to get it through DVD Architect. I created a Blu-ray ISO in DVD Architect with PCM audio. I mounted the ISO and extracted the 0000.m2ts file. Putting the 0000.m2ts file back on the Vegas Pro 13-453 timeline, I created the BMP files for the tool using all three Vegas De-interlace Methods; None, Blend and Interpolate. These are the results:


Date: 2019/01/13  12:40:16 
Description: Mainconcept AVC Blu-ray 1920x1080-60i Vegas De-interlace None 25-39 Mbps video stream
Frames Processed: 899
Processing Speed: 2.89 fps
Mean Squared Error: 12828.414
Peak Signal to Noise Ratio: 8.715

Date: 2019/01/13  13:31:32 
Description: Mainconcept AVC Blu-ray 1920x1080-60i Vegas De-interlace Blend 25-39 Mbps video stream
Frames Processed: 899
Processing Speed: 3.53 fps
Mean Squared Error: 8.086
Peak Signal to Noise Ratio: 43.594

Date: 2019/01/13  15:59:52 
Description: Mainconcept AVC Blu-ray 1920x1080-60i Vegas De-interlace Interpolate 25-39 Mbps video stream
Frames Processed: 899
Processing Speed: 3.58 fps
Mean Squared Error: 16.426
Peak Signal to Noise Ratio: 41.292

These are macro measurements of how faithfully the progressive file at the end of this process matches FHD series of bit maps rendered from the original progressive source. There are no subjective responses with regard to smoothness vs. sharpness.

Is your intent to use a dispassionate tool to decide which is "best" or collect the aggregate of the learned observations? a.k.a opinions

PS

When I was a traveling musician, I used to like to go to Pawn Shops in cities where I played to see the kinds of things people used in the location. Luckily, I never bought any of it (because I'd have to haul it around with me) and I never had to pawn anything. 

Musicvid wrote on 1/13/2019, 6:58 PM

Is your intent to use a dispassionate tool to decide which is "best" or collect the aggregate of the learned observations? a.k.a opinions.

Well, if being asked to choose, I'd have to say the latter; but my intention was more to offer a preliminary objectified reference around which to stimulate a discussion about everything else. To me, one deinterlacer looks better on my monitor, another better on my 50", and on my phone, well...nvm...

3POINT wrote on 1/14/2019, 4:02 AM

I preferred to leave interlaced, interlaced instead of deinterlacing. Nowadays I work only with progressive media.

Musicvid wrote on 1/14/2019, 9:46 AM

It's for people who convert interlaced to progressive for delivery.

3POINT wrote on 1/14/2019, 1:36 PM

It's for people who convert interlaced to progressive for delivery. Lots of people.

  • Delivery to what, YouTube? De-interlacing is always lossy, rather which method you use maybe not always less detail but most time it's less smoother.
  • There are some good de-interlacing methods as you mentioned, comparing them in a challenge, for me waste of time. When you want good progressive quality, start recording progressive.
Musicvid wrote on 1/14/2019, 2:14 PM

Thanks for sharing your thoughts. All interlaced material gets deinterlaced for all delivery (viewing). The inquiry into software solutions may not be relative to your interests, and I understand that completely, so no need to belabor.

That said, your posts have inspired and challenged me over the past year or so, in particular. Thanks for all you've contributed to these discussions.

3POINT wrote on 1/14/2019, 3:24 PM

your posts have inspired and challenged me over the past year or so, in particular. Thanks for all you've contributed to these discussions.

Should I interprete this positive or negative?

vkmast wrote on 1/14/2019, 3:30 PM

@3POINT, as positive, positively 😎

Musicvid wrote on 1/14/2019, 5:18 PM

Positive.

Former user wrote on 1/14/2019, 5:57 PM

@john_dennis

Hi John, thats a nice bit of lateral thinking indeed.

In VP16 theres a third gpu assisted deinterlace method also.

I had a go at doing a similar test but its just not right, i’m getting a huge mse value, 73, plus I can see differences when I use the difference compositing method. i.e. I rendered out the .m2ts file with blend mode to say blend.mov and compared blend.mov, now progressive, to the source progressive clip using the compositing difference method. Ok, I fixed this by setting deinterlace method to none which I had left on. This reduced the mse to 29, still just wrong.

 

Can you confirm what’s wrong in this ...

1 .. I selected a “source” origional FHD 25fps 27s clip, same data as my “General purpose” testing, progressive.  

2 .. Rendered out from “source” 675 .bmp frames to HORQ “Origional” folder for testing.

3 .. Rendered out “source” to .iso using tools/burn disc/blue ray disc.  Selected render image only.

4 .. Selected matching fhd 50i MC Mpeg2.

5 .. Extracted the 00000.m2ts file from the .iso file.

6 .. Loaded the 00000.m2ts file into VP, set deinterlace method to blend and rendered out the .bmp frames to HORQ “Rendered” folder for testing.

 

error at stage 3 .. failed to connect to an ipc port: the system cannot find the specified file.  So I used VP15 to create the iso.  This is a VP16 bug.  Not relevant to this query though.

Musicvid wrote on 1/14/2019, 8:24 PM

Note: If you downloaded the test file above, please do so again. The source image on the left is greatly improved. Test images on the right are unchanged. Toss this up on your big screen teevee!

Download the Test file mp4:

https://drive.google.com/file/d/1zB_oEF0KCwV0RT3kfhbv-grIiaBzrKlK/view?usp=sharing

@Former user Thanks for a great analysis.

john_dennis wrote on 1/14/2019, 9:00 PM

"Can you confirm what’s wrong in this ...

1 .. I selected a “source” original FHD 25fps 27s clip, same data as my “General purpose” testing, progressive.

       Good  

2 .. Rendered out from “source” 675 .bmp frames to HORQ “Original” folder for testing.

       Good

3 .. Rendered out “source” to .iso using tools/burn disc/blue ray disc.  Selected render image only.

      I didn't think of creating the ISO from the Vegas timeline. I rendered the interlaced Blu-ray video elementary streams (.avc) and separate (.wav) files to pass through DVD Architect to multiplex video and audio. I included video and audio sync puIses and beeps to check for audio drift.  I just realized for picture quality measurement, this isn't even necessary. Just put the interlaced .avc file back on the Vegas timeline and render to BMP image sequence in the HO test folder and skip to 7.     

4 .. Selected matching fhd 50i MC Mpeg2.

5 .. Extracted the 00000.m2ts file from the .iso file.

6 .. Loaded the 00000.m2ts file into VP, set de-interlace method to blend and rendered out the .bmp frames to HORQ “Rendered” folder for testing.

7. Run PQ Measurement.

Musicvid wrote on 1/14/2019, 9:22 PM

Coming soon -- numbers on my five samples from Otter stats.

 

Former user wrote on 1/15/2019, 6:06 AM

Thanks @john_dennis I give up now. I’ve spent considerable time, hours, and still cannot get it right. I did items 1 and 2. I then changed item 3 to this .. I rendered out source progressive to interlaced mpeg2. I then rendered out the interlaced mpeg2 file to an Xavc-i progressive file and then made the .bmp “rendered” stills from this progressive file. I then ran HO app. No .iso involved. Just for the record the figures are ...

Blend MSE / 60.282 PSNR / 31.291

Interpolate MSE / 66.384 PSNR / 31.441

So the gammy figures show the trend towards Blend, but its basically a rubbish test on my part.

So to summarise the roadblock I encountered... I rendered out to an upper field first Xavc-I clip, put it above the source progressive clip on the timeline and using the differencing test, established that it was quite low and should give a good mse and psnr result. In project properties field order had to be set to UFF. The deinterlace method, on or off had no effect.

What I found was that when using the HO app I could only get this low and probably correct value once I didn’t enable any deinterlace method in the project properties when rendering out the “rendered” sample .bmp's.

So of course I couldn’t compare the 3 different methods realistically.

Musicvid wrote on 1/15/2019, 2:29 PM

Added Edge Detail Comparisons for Five Deinterlacers:

Download PNG Here (1200x1400) Worth It:

https://drive.google.com/file/d/1Q_5Uj1vSX0JAmu21kzAzMV-cFIo5QcvL/view?usp=sharing

Former user wrote on 1/15/2019, 6:24 PM

The HO one looks the best?

Musicvid wrote on 1/15/2019, 6:40 PM

The HO one looks the best?

That looks like a job for Happy Otter RQMT. Probably only PSNR will mean anything because of varying render frame offsets

Yes, HO deinterlacer still looks the best to me too, but it sure doesn't play the best on my 50" Vizio, with my preferred movie viewing settings.

That's why the edge comparison alone is a joker.

My personal preference in my own living room, with my own refreshments, with my old eyes and a quite modest home entertainment system, are a tossup between HB Decomb and Yadif. Yadif plays a bit softer, while QTGMC rotation looks a bit stark. They all look about the same on my PC laptop.

Really, that's just me.

3POINT wrote on 1/16/2019, 2:10 AM

I'm still wondering, doesn"t play your 50" Vizio interlaced video?

Former user wrote on 1/16/2019, 4:00 AM

@Musicvid Would have been useful to see the Magix default, Blend, also Smart adaptive.

matthias-krutz wrote on 1/16/2019, 10:24 AM

Thank you for sharing the test file.
I did a test with the 60i source file and compared the Vegas native deinterlace methods. I rendered to 60p with resampling enabled.

https://1drv.ms/v/s!AqT8_NAXc5TbgXb0m6r29X50RSLA

For subjective checking I used my Sony projector. For me, Interpolate fields and Smart adaptive do a good job for this kind of material.
Please note: Dyn. RAM should be set to 0, when using the deinterlace Smart adaptive (GPU only) together with resampling. Otherwise render or timeline-playback issues are possible.

Musicvid wrote on 1/16/2019, 10:41 AM

 

I did a test with the 60i source file and compared the Vegas native deinterlace methods. I rendered to 60p with resampling enabled.

@matthias-kruz

You should render to 29.97p with resampling disabled.

That test, however is useful, and thanks for sharing your results!

I've thought about it, and testing "Bob" Deinterlacing at 60p would best be fodder for a different discussion.

 

 

Musicvid wrote on 1/16/2019, 1:44 PM

So, curiousity got the best of me.

@matthias-krutz, to put your Smart Adaptive 60p test on a level playing field, I placed it alongside Handbrake Decomb Bob 60p, where it fared very well indeed!

Going forward, we need to acknowledge that 60p will always play back "better" in a 60p environment than 30p, and so my wish to keep "Bob" in a separate discussion, lest we equivocate apples with oranges.

Musicvid wrote on 1/17/2019, 3:58 PM

Someone was going to ask for this eventually -- Source vs. Youtube deinterlace. Looks regrettably similar to Blend to these eyes, and plays just as soft. Like WMP.

Musicvid wrote on 11/18/2019, 7:10 PM

I have added edge test comparison for EEDI2 (HB).

I also retested it alongside HB Decomb Default, and the timed results are the same as they were in 2015 -- EEID2 takes conservatively 10x longer to render.