Infognition forum
October 24, 2017, 05:27:55 AM *
Welcome, Guest. Please login or register.

Login with username, password and session length
News: Last Video Enhancer version: 2.2
 
   Home   Help Search Login Register  
Pages: [1] 2
  Print  
Author Topic: Please update comparision :)  (Read 15796 times)
CruNcher
Newbie
*

Karma: +0/-0
Posts: 5


View Profile
« on: April 22, 2010, 09:51:05 PM »

I made myself a small one and this with the so high praised MotionDSP consumer Vreveal algorithm and im baffled how bad it is visually see the results bellow, there is so much bluring and temporal smoothing going on all the details get lost though it kills this way almost every artifact visible and so reaches a very high PSNR @ re compression and also low bitrate but the price you pay for this is to high imho.

http://mirror05.x264.nl/CruNcher/subjective_evaluation/SR/

PS:It will be interesting to see your version as a GPU accelerated one , also in comparison with consumer oriented SR like Arcsofts SimHD, MotionDSP Vreveal and Nero SR in the future.

Are you using the same algorithm as in the YuvSoft Super Resolution version in Video Enhancer ?

Even Neros SR seems to be visually better then MotionDSP, i mean when you keep the extra blur from the what seems to be internal TV scale conversion ;P

http://forum.doom9.org/showthread.php?p=1393943

made it a community thing Smiley



« Last Edit: April 23, 2010, 01:36:43 AM by CruNcher » Logged
Dee Mon
Administrator
Hero Member
*****

Karma: +13/-0
Posts: 743



View Profile WWW
« Reply #1 on: April 24, 2010, 12:35:58 PM »

Thanks!
By the way, how did you make that MotionDSP result? It seems that in vReveal there are at least two different ways to perform an upsize - one strictly 2x and another almost arbitrary, and the latter seems to be plain resampling, don't know about the former.

We're planning to make a new comparison with all the new competitors and also AviSynth scripts mentioned on doom9. The old comparison is a bit outdated now. BTW you guys need to use another image hosting since most images in that thread are not showing up, instead I see messages about hotlinking.

YUVSoft SR is partially based on our method (I personally participated in developing it a few years ago) but it evolved further, providing higher quality at the cost of speed. Afaik it's really slow now.
Logged
MC87
Newbie
*

Karma: +0/-0
Posts: 23


View Profile
« Reply #2 on: May 21, 2010, 06:32:07 AM »

MotionDSP's Vreveal 2x resize feature uses Super Resolution (though it's only available on videos 288p or less).

Theoretically their resize should actually better than Video Enhancer's as apparently it uses both earlier and later frames to form the final image, whereas VE only uses earlier frames. And MotionDSP is a high end provider of such enhancement software to the military, police etc.

I'm not sure though, but are you sure you used the 2x resize feature or just their regular simple resize which doesnt use Super Resolution?
Logged
Manoman455
Full Member
***

Karma: +4/-0
Posts: 109


View Profile
« Reply #3 on: May 23, 2010, 07:54:45 AM »

Is this guy serious? i have used both software and VE wins hands down, Vreveal cant even load in as many different videos that VE can. Also their other enhancements suck. And im not sure but i thought VE uses frames from both previous and futures fames to recreate the current frame.
Logged
Dee Mon
Administrator
Hero Member
*****

Karma: +13/-0
Posts: 743



View Profile WWW
« Reply #4 on: December 08, 2010, 07:10:03 PM »

Finally the new comparison is ready!
Tools and methods compared:
3 VirtualDub resize methods (bicubic, Lanczos3, Lanczos3+sharpen),
3 AviSynth plugins (Spline36, EEDI3, NNEDI3),
2 plugins for After Effects (InstantHD, Boris FX Uprez),
1 Photoshop plugin (Genuine Fractals),
5 standalone applications: CinemaHD, vReveal, Video Enhancer, SAR Image Processor (4 different algorithms), SmartEdge.

http://www.infognition.com/articles/video_resize_shootout.html

17 methods compared on 10 video files and 10 single frames, PSNR objective metric measured and charted, 304 sample images included.
Logged
MC87
Newbie
*

Karma: +0/-0
Posts: 23


View Profile
« Reply #5 on: December 11, 2010, 05:20:23 AM »

Great comparison update!

vReveal does blur the IQ too much - It's a pity that their consumer product is so watered down compared to their high end forensic tool Ikena (which is very powerful if you look at the website)

For the Video Enhancer results, did you use 1 application of Super Resolution or 2 chained filters? If you used two, how much better PSNR is gained versus using only a single resize?

Also have you got any PSNR stats for Video Enhancer where both preceding and succeding frames are used in the Super Resolution algorithm?

« Last Edit: December 11, 2010, 05:27:17 AM by MC87 » Logged
Dee Mon
Administrator
Hero Member
*****

Karma: +13/-0
Posts: 743



View Profile WWW
« Reply #6 on: December 11, 2010, 01:31:19 PM »

>Ikena (which is very powerful if you look at the website)

There is a saying: every program works perfectly in PowerPoint. The same applies to websites with descriptions of products you cannot try. Wink

Video Enhancer results were obtained with a chain of two SR filters.
We'll measure results of single SR application and report here after this weekend.

No, only forward-direction processing was measured. Version with bidirectional processing is only planned for future releases, it's not done yet.
Logged
Dee Mon
Administrator
Hero Member
*****

Karma: +13/-0
Posts: 743



View Profile WWW
« Reply #7 on: December 14, 2010, 11:49:52 AM »

We measured the same files upsized with one instance of SR. The results are just 0.06-0.17 dB lower in PSNR, which means in many cases one pass is good enough.
http://www.infognition.com/blog/2010/resize_video_4x.html
Logged
Henery
Newbie
*

Karma: +4/-0
Posts: 35


View Profile
« Reply #8 on: December 14, 2010, 01:34:00 PM »

PSNR losses that small are quite good news.With only one SR filter applied,it should reduce processing times.
Anyway,i have two ideas for where super-resolution or the high-quality image resampling method could be used:

1)Photoshop image enlarging plugin.I think it would still beat even Genuine Fractals without multiframe analysis.
2)Realtime Youtube video upscaler.480p --> 1080p enlargement shouldnt take so long with todays computers,especially with one SR filter chain applied.I will test this today if i have time.

Cyberlink has made a youtube video enhancement plugin working in realtime but it has quite poor upscaling.
Logged
tritical
Newbie
*

Karma: +0/-0
Posts: 2


View Profile
« Reply #9 on: December 15, 2010, 08:08:11 AM »

Interesting comparison. Any idea when the high resolution originals will be up? I'd like to try some tests and to duplicate some of the results. Also, what downsizing method was used?
Logged
Dee Mon
Administrator
Hero Member
*****

Karma: +13/-0
Posts: 743



View Profile WWW
« Reply #10 on: December 15, 2010, 09:03:01 AM »

Henery,
1) Our method is not applicable to single images (nowhere to get information from). NNEDI seems to be a good choice for single image enlarging.
2) In YV12 mode (native for youtube video) we get up to 20 FullHD (1080p) frames per second on a quad core. This is still not fast enough for realtime. And most PCs and laptops now do not have 4 cores yet...

I've seen that Cyberlink plugin a year ago. It worked only in Internet Explorer and the quality of resizing wasn't good (bicubic at best).

tritical,
The hi-res files are quite large (~300 MB each), we're uploading them to rapidshare now. Will be available in a couple of days.

Downsizing was made by averaging corresponding pixels, i.e. each 4x4 block was averaged to 1 pixel. This approach more closely resembles difference between hi-res and low-res matrices in cameras and provided higher quality than simple bilinear resize. I was going to make a separate blog post dedicated to this question.
« Last Edit: December 15, 2010, 09:06:51 AM by Dee Mon » Logged
Dee Mon
Administrator
Hero Member
*****

Karma: +13/-0
Posts: 743



View Profile WWW
« Reply #11 on: December 17, 2010, 07:49:34 AM »

Links to the high resolution original videos have been added to the comparison page.
Everyone is welcome to reproduce and check our results.
Logged
CruNcher
Newbie
*

Karma: +0/-0
Posts: 5


View Profile
« Reply #12 on: December 19, 2010, 02:53:48 PM »

Nice Dmitry Smiley and 100% my Vreveal Consumer experience (like extreme high FFT Sigma setting), though we should be fair its optimized for very low quality source and i believe them if they say they have something much better (they just don't use it in the consumer space) also according to their presentations it makes sense, i doubt US Air Force would use something of no quality for their Reconnaissance Data Analysis Smiley

MotionDSP tech talk (algorithm  & software) by one of the main Devs http://nvidia.fullviewmedia.com/GPU2009/1002-atherton-1459.html
MotionDSP executive presentation (usage field talks of Ikena) http://nvidia.fullviewmedia.com/gtc2010/0923-a5-2027.html

Their Ikena Live (Realtime GPU SR) Demo Video http://www.youtube.com/watch?v=Typ03BN8ARo

Somehow i miss http://www.yuvsoft.com/download/super-resolution/index.html in the comparison Smiley (your ex home) i guess we would see very close results just @ much lower speed as you already explained before.
I did a lot of test with the Smith and other scenes and Video Enhancer has a tendency to create heavy aliasing on the edges in the final result (when doing extreme upscales) compared to the other things i tried (also my own Avisynth GPU based framework) that rather like to blur to avoid this which in the end looks much better in motion.
So im not sure if its really a good way to just rate the Visual Quality with PSNR but you provide the screenshots and those doesn't hide this aliasing problems of Video Enhancer @ all, seeing those in Motion though is a entire different thing and can make Vreveals Super Blur look nice, even as crazy as it sounds.

But for Consumers no doubt Infognitions is one of the most advanced SR available Smiley
Logged
Dee Mon
Administrator
Hero Member
*****

Karma: +13/-0
Posts: 743



View Profile WWW
« Reply #13 on: December 19, 2010, 08:25:35 PM »

Thanks for the links!
We forgot to check Yuvsoft, thought it wasn't accessible for wide audience. We'll do the tests with its demo.

Yes, PSNR results do not reflect exactly human perception, this is just a simple similarity measure, but at least it's objective and well known in the field of computer graphics and digital video.

Seeing the Ikena clip on youtube, it confirms my thought that their "SR" just blends together motion-compensated frames, this reduces noise but doesn't really improve resolution. Other major things shown there can be done in Video Enhancer or VirtualDub using auto-levels and Deshaker filters. Smiley
Bulding a panorama out of video is a different task, I remember MSU lab had some working example of this.
« Last Edit: December 19, 2010, 08:35:18 PM by Dee Mon » Logged
MC87
Newbie
*

Karma: +0/-0
Posts: 23


View Profile
« Reply #14 on: December 20, 2010, 01:06:49 AM »

With vReveal, the problem is you are also making it resize to match the 2x or 4x (depending on how many SR filters are being chained) resolution output from VideoEnhancer.

vReveal doesn't upscale the frame during it's SR process - any upscaling is done after the SR and filters are applied and it is only done by a bicubic filter. This is one reason why it's output is so blurry.

To get a fair comparison, you should use a eg. 240p video and get vReveal to output at the same resolution as the input video (but tick the 'enhance with super resolution' box) and then you should do the same with Video Enhancer (ie. do not resize the input like it normally will) and then compare the PSNR values of the two.
Logged
Pages: [1] 2
  Print  
 
Jump to:  

Powered by MySQL Powered by PHP Powered by SMF 1.1.13 | SMF © 2006-2011, Simple Machines LLC Valid XHTML 1.0! Valid CSS!