|
|
| pookien |
| Posted: Jan 14 2014, 03:56 PM |
 |
|
Advanced Member
Group: Members
Posts: 39
Member No.: 24784
Joined: 30-December 08
|
It's an interesting theory. Let's test it in practice. |
|
| evropej |
| Posted: Jan 14 2014, 08:53 PM |
|
|
Advanced Member
Group: Members
Posts: 514
Member No.: 26523
Joined: 28-November 09
|
Pookien, are you the developer? |
|
| meowmeow |
| Posted: Jan 15 2014, 02:41 AM |
|
|
Advanced Member
Group: Members
Posts: 98
Member No.: 37160
Joined: 21-September 13
|
| QUOTE (pookien @ Jan 14 2014, 03:56 PM) | | Let's test it in practice. |
I would with SSIM or PSNR using AviSynth but this VirtualDub filter says
"Unregistered version is fully funcional but draws some watermark over video." |
|
| pookien |
| Posted: Jan 15 2014, 11:24 AM |
|
|
Advanced Member
Group: Members
Posts: 39
Member No.: 24784
Joined: 30-December 08
|
Are you expecting to see meaningful PSNR or SSIM results for that "psy" thing?
You can compare the videos visually first. When I open a 720x384 HDRip and apply this filter result is clearly different from player's default bicubic resampler. For example:
http://data.infognition.com/avssr/tf12323a-bili.png (MPC's bilinear)
http://data.infognition.com/avssr/tf12323bc.png (MPC's bicubic)
http://data.infognition.com/avssr/tf12323sr.png (SR)
You can upload your test video and I'll process it with SR so you could measure and compare PSNR/SSIM.
evropej,
actually yes, I am. |
|
| meowmeow |
| Posted: Jan 15 2014, 02:28 PM |
|
|
Advanced Member
Group: Members
Posts: 98
Member No.: 37160
Joined: 21-September 13
|
| QUOTE (pookien @ Jan 15 2014, 11:24 AM) | | Are you expecting to see |
original -> 1/2x res - > x264psy -> lanczos 2x res
original -> 1/2x res - > x264psy -> SR 2x res
original -> 1/2x res - > x264nopsy -> lanczos 2x res
original -> 1/2x res - > x264nopsy -> SR 2x res
original -> SSIM/PSNR -> lanczos psy/nopsy ->diff1
original -> SSIM/PSNR -> SR psy/nopsy ->diff2
diff1 <> diff2
| QUOTE (pookien @ Jan 15 2014, 11:24 AM) | | For example: |
I see sharpening with temporal denoise.
| QUOTE (pookien @ Jan 15 2014, 11:24 AM) | | You can upload your test video |
No, its very expensive process. |
|
| evropej |
| Posted: Jan 16 2014, 04:39 PM |
|
|
Advanced Member
Group: Members
Posts: 514
Member No.: 26523
Joined: 28-November 09
|
pookien,
You ever consider adding deblocking to your software?
The probability of low resolution videos being not compressed is less than 99% of the time which means they are blocky because of compression.
 |
|
| raffriff42 |
| Posted: Jan 16 2014, 09:25 PM |
|
|
Advanced Member
Group: Members
Posts: 384
Member No.: 35081
Joined: 25-June 12
|
I think you people are being a bit too hard on the developer here
1. No upscaler is good with noisy, blocky sources. Fix your source first.
2. For anime, use nnedi (edge-directed interpolation)
3. This filter uses a specialized algorithm that performs a neat trick - on specific source types. It is one upscaling tool among many. Ninety percent of the time, good old Bicubic works about as well as anything 
|
|
| evropej |
| Posted: Jan 16 2014, 09:40 PM |
|
|
Advanced Member
Group: Members
Posts: 514
Member No.: 26523
Joined: 28-November 09
|
Hey, at first I thought it was a guy posting a thread and didnt know he was the developer. Now with all the discussions, we almost did some advertising for him lol. If we never said anything, the post would have withered away
I still say that deblocking would be nice, most deblocking software is unstable and causing crashing.
I would also add that that the cost of the filter is high. If the filter did more and cost less, it would generate more interest.
Just my 2 cents. |
|
| pookien |
| Posted: Jan 17 2014, 04:46 AM |
|
|
Advanced Member
Group: Members
Posts: 39
Member No.: 24784
Joined: 30-December 08
|
Actually, in our previous version of SR engine used in Video Enhancer there was an optional deblocking pass for YV12 mode (pic). But it wasn't included in the new version for a number of reasons:
1. If your video is so much compressed that blocking is visible, then in many cases too much information is already lost during compression, and SR will not gain much.
2. Deblocking is often a part of a decoder, and a good decoder can do it better.
3. Speed: old SR engine was based on image-wise operations, current version is block-based which is good for CPU cache and makes it so much faster. For deblocking we need to make a new image first (with some changes on block borders and no changes inside them), that means requiring more memory and polluting more cache.
But if you think it's worth it we can add the deblocking back, we already have a working implementation (link). |
|
| evropej |
| Posted: Jan 17 2014, 03:23 PM |
|
|
Advanced Member
Group: Members
Posts: 514
Member No.: 26523
Joined: 28-November 09
|
From my experience, all video which is small in pixel dimensions, suffers from blocky compression.
Its just my suggestion based on videos I tried to upscale up with vdub.
So, with that said, let me say this:
Even if information is lost, when the eyes see a video, the blocks tend to hide available information. Removal of block will make the eyes perceive and increased detail level.
The same principle applies to noise removal. The eyes are not distracted by noise and hence focus on available detail.
Some people say that upscaling cant create information that is not there which is technically true. But, ask yourself this, why can you see so much detail on the moon when its on the horizon compared to when its overhead? Its the same size, I have proved this to myself by taking pictures at both positions and comparing them. So upscaling again make the eyes perceive more detail.
I made a comment about the cost of the converter because most of the people here are freeware users. Introducing something which costs money will most likely go unnoticed unless the price is under a threshold.
PS Just read your company profile. You just tied a knot for me, the connection with MSU. What happened to the MSU developers? Did they all branch off like you? |
|
| pookien |
| Posted: Jan 18 2014, 03:16 AM |
|
|
Advanced Member
Group: Members
Posts: 39
Member No.: 24784
Joined: 30-December 08
|
| QUOTE | | Some people say that upscaling cant create information that is not there which is technically true. |
It's true for usual intra-frame methods like bicubic or Lanczos. Some methods like fractal resize can make up some details that were not there originally and hope it looks ok. I.e. it's not restoring original information but adding some synthetic info. Super resolution is a bit different: it passes information from one frames to others, so while it technically cannot get any new information that wasn't present in the video as a whole, for each particular frame it can add some info not present in this frame but present in neighbor ones.
A simple illustration:
| CODE | "t*is*is*a *en*en*e *ith *om* in*o**at*on"
"*h*s *s*a s*n**n*e*wi*h *ome*inf*rm*t*o*"
"*h*s i* * *e*te*c* w*t* s**e **fo*m**i**"
"t*i***s **se***nce***th **me ***or***i*n" |
I think your brain could restore all the words even though no word is present fully in any line.
As for MSU VideoGroup, one needs to understand that this is a university group, where 98% of people are just students who spend some time there, make their theses and go away. And they also do commercial projects for big companies like Samsung and Intel, which take a lot of their time and efforts, leaving not much for the free filter development. |
|
| evropej |
| Posted: Jan 20 2014, 02:29 PM |
|
|
Advanced Member
Group: Members
Posts: 514
Member No.: 26523
Joined: 28-November 09
|
I understand but cant the professor assign the work to new students?
Or release the source code so other can patch them? |
|
| pookien |
| Posted: Jan 21 2014, 06:33 AM |
|
|
Advanced Member
Group: Members
Posts: 39
Member No.: 24784
Joined: 30-December 08
|
I haven't seen them for a while but I've heard quality of students dropped significantly in the last few years due to changed rules for entering the university (before each department held their own entry exams to filter good students, now they admit students just by their scores in global school graduation tests which doesn't work well; but I might be mistaken here).
And the source code, apart from licensing issues, is usually so horrible you wouldn't want to maintain it.  |
|
| evropej |
| Posted: Jan 21 2014, 02:43 PM |
|
|
Advanced Member
Group: Members
Posts: 514
Member No.: 26523
Joined: 28-November 09
|
The same problem exists here, they are pushing people into Universities who dont belong there and cause the general education level do drop. Getting a 4 year degree used to be a good education, now you need to go 6 years to get the same education. For what we call highschool or middle school ( grade 7-12 ), you can even fail every class and pass. Yes, I am not making this stuff up, its called no child shall be left behind law.
As for the MSU code, what kind of licensing is applicable? Why not open the source code for personal use. I am sure people will add value to it and at least fix some of the major issues with it like crashing lol. Just an idea. |
|
| pookien |
| Posted: Jan 23 2014, 07:43 AM |
|
|
Advanced Member
Group: Members
Posts: 39
Member No.: 24784
Joined: 30-December 08
|
As I understand some of the filters evolved into code which was sold to Samsung etc. If those companies got exclusive rights for the evolved code, open sourcing earlier version of the code might be a complicated legal issue. Although usually MSU VideoGroup tried to avoid exclusive licensing, preferring non-exclusive. |
|
