AVS Forum banner

Status
Not open for further replies.
61 - 80 of 111 Posts

·
Registered
Joined
·
1,358 Posts
Quote:
Originally posted by NiToNi
What kind of issues Vlad?


It shows as blocking, almost like a grid, in my system.


How many bits precision today?


Don't know, but would doubt it's more than 8-bit.


What is ffdshow Lanczos today, Catmull-Rom or Spline? Obviously, we need Catmull-Rom, but is Andy really the right guy to develop this? Integrate the code, yes, but to write the algo itself?


Not sure again. Andy was asking in another thread for some guidance where could he find better algos as he wanted to integrate that better algo with 32 bit floating precision into a new algo.


What is aSharp really doing to the image, in technical terms, especially at the settings you prefer to run it at?


Except for video card/driver issues, I am more practical than theory guy. At one point I did some read on it, but that was a while back. There was a lot of discussion on Asharp on Avisynth form.


Have you played with LimitedSharpen as suggested by Socio?


Not really. I am home only for weekends for the last couple of month, so just don't have time. At this point I am quite happy with the image quality and Asharp does a great job sharpening without excessive ringing, when set up correctly. The key is finding the right balance between unsharp masking threshold and adaptive sharpening strength. The two are inter-dependent. I'd reduce adaptive sharpening strenght to the point you can tolerate ringing at certain level of unsharp masking threshold (probably between 0.1 and 0.8), and then play around with more combinations.
 

·
Registered
Joined
·
1,563 Posts
Doesn't the June 7th version of ffdshow have lanczos2 ?

I mean parameter 2 does mean lanczos2 right?

:confused:
 

·
Registered
Joined
·
7,054 Posts
Quote:
Originally posted by Nima
Doesn't the June 7th version of ffdshow have lanczos2 ?

I mean parameter 2 does mean lanczos2 right?

:confused:
That's what I thought too.
 

·
Registered
Joined
·
341 Posts
Here are two sets of 960x1440 screen captures that show the result of different sharpening schemes using ffdshow August 2, 2004 SSE build.


The filters used were Picture Properties (gamma 1.01) and Resize. Also Avisynth histogram was used. Output was YV12. Each file is nearly 1MB which is as small as non-lossy compression will produce. Sorry about the size but lossy compression adds artifacts to he images which makes them useless. The the images are best viewed full screen to eliminate the light flooding from the browser window. Just copy the images to Photoshop or whatever you use and it makes it easy to compare them.


The images were chosen to illustrate any image deterioration in areas of medium to low contrast. The reference images are the ones using the Simple Resize filter since it does minimal signal processing.


Simple Resize:
http://acme99.dyndns.org/E5LunchWago...d040802SSE.png


Bicubic Resize with luma sharpen at 0.0:
http://acme99.dyndns.org/E5LunchWago...d040802SSE.png


Bicubic Resize with luma sharpen at 0.5:
http://acme99.dyndns.org/E5LunchWago...d040802SSE.png


Lanczos Resize parameter at 2 and luma sharpen at 0.0:
http://acme99.dyndns.org/E5LunchWago...d040802SSE.png


Lanczos Resize parameter at 2 and luma sharpen at 0.5:
http://acme99.dyndns.org/E5LunchWago...d040802SSE.png


Lanczos Resize parameter at 4 and luma sharpen at 0.0:
http://acme99.dyndns.org/E5LunchWago...d040802SSE.png


Lanczos Resize parameter at 4 and luma sharpen at 0.5:
http://acme99.dyndns.org/E5LunchWago...d040802SSE.png


Simple Resize:
http://acme99.dyndns.org/E5Tear 35-3...d040802SSE.png


Bicubic Resize with luma sharpen at 0.0:
http://acme99.dyndns.org/E5Tear 35-3...d040802SSE.png


Bicubic Resize with luma sharpen at 0.5:
http://acme99.dyndns.org/E5Tear 35-3...d040802SSE.png


Lanczos Resize parameter at 2 and luma sharpen at 0.0:
http://acme99.dyndns.org/E5Tear 35-3...d040802SSE.png


Lanczos Resize parameter at 2 and luma sharpen at 0.5:
http://acme99.dyndns.org/E5Tear 35-3...d040802SSE.png


Lanczos Resize parameter at 4 and luma sharpen at 0.0:
http://acme99.dyndns.org/E5Tear 35-3...d040802SSE.png


Lanczos Resize parameter at 4 and luma sharpen at 0.5:
http://acme99.dyndns.org/E5Tear 35-3...d040802SSE.png


Charlie
 

·
Registered
Joined
·
2,690 Posts
Very nice screen shots Charlie. Perhaps you could label them in your post for easy reference: 1a, 1b, 1c..., 2a, 2b, 2c...


I only looked at the 3rd, 5th, and 7th shots of the 2nd picture but could see no differences at all in PQ.
 

·
Registered
Joined
·
971 Posts
It's more difficult to see the differences in the face shots because there are no hard, straight edges. You have to look closely, but the diiferences can most easily be seen in her hair. You can see the differences in the face itself, but you kind of have to stand back from it a bit.


On the wagon pictures, it is much easier to see. Especially if you toggle betwen the simple resize and the Lanczos4. The difference between the L4 and L2 is very slight. If I were comparing them sided by side instead of toggling the picture in Irfanview, I might not even be able to see the difference.


RG
 

·
Registered
Joined
·
5,344 Posts
my CPU sure likes Lanczos 2 !!!!


:D
 

·
Registered
Joined
·
341 Posts
1Time,


Would it help if I put the parameters at the top of each image near the OSD information? The file names have the parameters too but get lost in full screen mode.


Differences between images are subtle and may be masked by stray light or large spot size.


Charlie
 

·
Registered
Joined
·
9,884 Posts
To put things into perspective for the comparison, SimpleResize is just a pure unfiltered linear interpolation between 2 adjacent pixels, optimized mostly for speed.


I wrote it for downsizing the increasingly common HD caps that already appear to be excessively filtered or soft for some reason and thus hopefully need no other filtering before (re)compression to Xvid.


SimpleResize really doesn't upscale all that well.


- Tom
 

·
Registered
Joined
·
317 Posts
Thanks Charlie. The photos are very helpful to visualize what's happening with each setting change.


Jason
 

·
Registered
Joined
·
341 Posts
Tom,


When you say that Simple Resize doesn't upscale well I assume you are referring to it's soft image. I can't find fault in the way it upscales to 960x1440 with that one caveat. It has become important in my video evaluation as a reality check to gage how much quality I'm sacrificing for sharpness. I'm sure glad you wrote it and made it available.


Thanks,


Charlie
 

·
Registered
Joined
·
2,690 Posts
Quote:
Originally posted by Charles Black
1Time,


I put the parameters in each image.


Charlie
Much better, thanks again Charlie.


I just compared the 3rd and 5th shots of the 1st picture and I can't find a difference, and I'm a fairly detail oriented guy.


So which one looks the best to anyone else?


Also, would SSE2 would be "that" much better than SSE when using FFDShow?
 

·
Registered
Joined
·
1,638 Posts
I agree 1Time, thats a pretty tough call. Maybe I'll look at them on my PJ. A big screen sometimes shows things I can't see on a 17" monitor. Hell, maybe its time to go back to bicubic :p
 

·
Registered
Joined
·
2,690 Posts
Right, it's time to check these out on the big screen and get back to fiddling with FFDShow again. I just haven't got it right yet.
 

·
Registered
Joined
·
341 Posts
1Time,


The Lanczos p2 ls0.5 (fifth shot) has more sharpening - the finest details are effected most. Look at the buildings in the distance.


I just noticed that two of my examples are not the same frame so back to the drawing board. Edit: Their now fixed.


Charlie
 

·
Registered
Joined
·
3,314 Posts
Just to keep some people happy.


Does ffdshow linearise the video material prior to scaling ie does it attempt to flatten the video intensity curve into a totally linear response ( a colour correction or transfer function if you will).


I'm not sure how much of an impact the nonlinearity of a video curve has on pixel interpolation but its certainly common place to linearise 10bit log film scans prior to filtering operations to get more "accurate" pixel values in the interpolation.
 

·
Registered
Joined
·
341 Posts
Mr.D,


I have checked the ffdshow luma output exhaustively with a scope and test signals and it is linear as expected. I checked chroma ranges and they were as expected as well. My chroma experience is limited to a few samples though and would not have been able to see any chroma errors other than range.


Charlie
 

·
Registered
Joined
·
2,164 Posts
A good way to make comparisons (without resorting to multiple layers in photoshop) is to open each image in a separate TAB in Firefox, then switch between them, if you do this, you will notice the difference bettween images 3 and 5 and 7 much more easily with the first scene.


Cheers...

Duy-Khang Hoang
 

·
Registered
Joined
·
2,690 Posts
Charlie, thanks again for this comparison.


I now can see the differences. I put the 3rd, 5th, and 7th shots of the 1st picture in IrfanView and flipped through them. For me it came down to the definition and darkness of faint, thin lines on a few buildings in the background.


3rd = Bicubic, sharpen 0.5

5th = Lanczos 2, sharpen 0.5

7th = Lanczos 4, sharpen 0.5


The 7th shot was the best followed very closely by the 5th shot and the 3rd shot was easily the worst, that is, when I focused my attention on these faint, thin lines in the background.


On a scale of 1 to 10 and rating the 7th shot as a 10, I'd rank the 5th shot as an 9 and the 3rd shot as a 5.


So from what I can see on my monitor there's not that much of a difference between Lanczos 2 and 4. So I'll be giving Lanczos 2 a try with my pj since my 2500+ Barton (overclocked to 3200+ speed) and my ATI 9700pro have been struggling with my latest FDDShow settings. I may not upgrade my processor and video card so soon after all.
 
61 - 80 of 111 Posts
Status
Not open for further replies.
Top