or Connect
AVS › AVS Forum › Video Components › Home Theater Computers › ffdshow Resize filter to Sharpen
New Posts  All Forums:Forum Nav:

ffdshow Resize filter to Sharpen  

post #1 of 168
Thread Starter 
I have not seen anyone discussing the use of ffdshows Resize filter to sharpen.
I find that using the Luma and Croma sharpen function in ffdshow resize filter gives a much more natural effect than any other sharpen filters IMHO.

Also the order the filters are used is important for performance and quality.

My recommendation for filter run ORDER in ffdshow is as follows.

1. Deinterlace (If Required)
2. Levels and or Picture Properties
3. Descaler (GradualNoise) or Blur & NR (Gradual Denoise or Temporal Smooth)
GradualNoise is a CPU Killer. Temporal Smooth is less so. Gradual Denoise is easy on the CPU and is my recommendation.
4. Resize (Bicubic) up to at LEAST 25-30% higher than DVD resolution. 200% is good, but hard on CPU. Do this even if your display is only low resolution.
Let the video card rescale the picture back down for output to display.
This gives a very subtle smoothing effect.
5. Last, us the Luma and Croma sharpening in the resize filter.
This looks more natural IMHO than any other shapening filters I have tried.
The Resize and Sharpening needs lots of CPU power.

If you have a fast CPU you can try running the Denoise filters last, after the resize filter so that it is working on the upscaled image. This can look really good but it is taxing on CPU.

NOTE: You can change the order of filters by using the up/down arrows to the left of a selected ffdshow filter.

So people. Give this setup a go and let me know what you think.

Regards,

Owen.
post #2 of 168
I read your post in the other thread and gave it a go last night.

I don't think anybody has played with resize, I didn't really notice it before - for some reson it doesn't have a check box on the lhs of the fddshow screen. I also didn't realise that you could change the order of the filters!

I have a 2200+ athlon XP, but Blight is right, the resize is brutal!! I don't have enough CPU to use it. To check the sharpening in resize settings I ended up resizing from 720x576 to 720x576, but I couldn't see any real changes. Maybe it doesn't work with no actual size change...

I also found that resize does not change on the fly, you have to stop and start playback. This makes it difficult to make quick comparisons.

However playing around with the filter order is interesting, it does make a difference.

Also, you mention that temporal smooth is easy on the CPU - I have found the exact opposite (and other early fddshow threads found the same), I remember reading a post from Blight (I think, maybe Ogo), saying that temporal smooth is really for encoding, not for real time playback.

Then I started playing with the dscaler noise reduction filters - haven't really sorted them out yet, so may choices.

The problem with fddshow + dscaler is that there are 3 ways to sharpen, and about 2000 ways to remove noise, its a little complicated......

VMR9 also gives me a huge CPU hit as well, I wonder if the video card affects the amount of CPU required to use it? I have a cheap & crappy Geforce II MX400.

Might try the older fddshow too...

Just when I thought I had fddshow sorted - you went and threw a cat amongst the pigeons!

Mark
post #3 of 168
tried resize and I agree that it sharpens nicely. However my cpu use jumps to 100% and I start to lose frames. I'm sticking with asharp, have you guys tried that? At a threshold of 1.07 and a strength of around 4-5 it looks pretty nice while avoiding SOME of the artifacts that regular unsharpmask produces.
post #4 of 168
Thread Starter 
Mark_A-W

Thanks for bringing my attention to my oops regarding Temporal Smooth filter. :D
I have edited my above post to correct same.

Your Geforce 2 MX is probably to slow for VMR9.
When I purchased my Geforce3 Ti500 for my HTPC to replace a Geforce 2 MX400, it was only to get the best Svideo TV out. (New Conexant CX25871 chip only available on GF3 and GF4 at the time).
I dont play games so I thought the fast card was a waste.
Well now I'm very glad of the speed to run VMR9.
Anyone who thinks that a slower Video card will do, is now mistaken.
Radions have a nice overlay so VMR9 is not required.

I should bring your attention to point 4.

4. Resize (Bicubic) up to 1024-768 or MORE even if your display is of lower resolution.

Dont expect good results if cant do this due to CPU constraints.
Use overlay mode in stead of VMR9
Then use the sharpening method I suggested. In the resize filter.
For noise reduction I like Descaler (GradualNoise) but its VERY hard on CPU.
I find that Cinemaster decoders work best for me. 100% CPU usage with the above filters and Reclock but smooth playback.
NVDVD decoders give only 80-90% CPU but playback stutters.

Regards,

Owen

P.S. Where can I get a 4 Gig P4 ??

:D
post #5 of 168
Is there documentation anywhere about what constrains performance of specific filters? As in what filters bottleneck on memory-bandwidth and which ones bottleneck on cpu speed. It would also be useful to know which filters work better on a P4 (use SSE2 probably) and which filters work better on the Athlon (most likely using the older x87 floating point instruction set). This kind of information would be useful in determining what sort of system to buy as well as what approaches (if any) towards over-clocking might yield good results.

For example a filter that uses SSE2 and is memory-bandwidth constrained would probably benefit from a P4 with 553MHz bus (or even wait a month for the 800MHz bus) while a cpu-constrained filter that doesn't use SSE2 might better off on an Athlong system.
post #6 of 168
Few points:
1. There's no point resizing unless you resize upward at least 20-30%. The idea behind upsizing is to give the sharpening filters a bigger work area.
2. I don't think VMR9 should be such a CPU hit. DO NOT use the NVIDIA 4x.xx drivers, I find the 30.82 driver to be the most stable for video playback. The newer drivers seem to have "issues".

As far as which filters give a big CPU hit. Basically any that are either unoptimize, or that have a massive number of calculation to do. Warpsharp/ASharp/Resize are pretty heavy. Frankly, if you have an HD display (1400x[n] or higher), you may want to start looking at a multi-CPU system for the ultimate in high-end processing. A quad P4 3ghz system could do wonders on the image (and your wallet).
post #7 of 168
Thread Starter 
Blight,

Quote:
1. There's no point resizing unless you resize upward at least 20-30%. The idea behind upsizing is to give the sharpening filters a bigger work area.
Absolutely, this cant be stressed enough.

Quote:
2. I don't think VMR9 should be such a CPU hit. DO NOT use the NVIDIA 4x.xx drivers, I find the 30.82 driver to be the most stable for video playback. The newer drivers seem to have "issues".
I am using an old Driver 6.13.10.xxxx should I upgrade??

CPU usage as shown in Task Manager is not that much higher but the ability to run smoothly sure can be.
It's probably just the straw that brakes the Camels back.
The same applies to filters. I find that Descaler (GradualNoise) is most difficult to run smoothly followed by ffdshow Temporal Smooth and then ffdshow GradualNoise.
I have no explanation for this, it’s just my observation.
Its good to clear up these points for others reading here.

Hey you forgot the :D at the end of your post RE: the Quad P4 :D

Regards,

Owen
post #8 of 168
Owen, you should probably give the 30.82 a shot and if it doesn't work better for you, revert.
post #9 of 168
VMR9 forces the DVD filters into software mode, this is where a lot of the increased CPU is coming from.

Vern
post #10 of 168
VMR9 forces filters into Direct3D mode. Scaling should be done by the D3D texture unit. iDCT and Motion Compensation should still be supported through DXVA.

This depends on the decoder filters and display drivers properly supporting these...
post #11 of 168
Thread Starter 
Come on people,
How many of you have tried the software Resize with Sharpen/Soften filter in ffdshow?
For me the quality improvement over hardware resize is very noticeable.
Is this because I'm using a Geforce card?

Maybe Radion hardware resize is much better than Geforce, I know
Overlay is better. That’s why I use VMR9 on the GF.

I am resizing standard PAL 720-576 to 1024-768 or 1280-1024
I run a Desktop Res of 1024-768 which is output to a calibrated Toshiba
43" Non HD rear projector through a good quality Svideo cable using TVTool.
The TV out chip on my GF3 Ti500 (CX25871) down samples my desktop to PAL 768-576 for output to TV.
So now I have a 1024-768 Desktop showing on the 768-576 TV screen.
The visible (non overscaned) resolution is 720-540i. Not to bad.
I do not have VGA input on the TV. Only Svideo and Component.
Even so, my 1024-768 Desktop is quite sharp and readable from 3 meters viewing distance.
Surprisingly DVD video quality looks better than outputing a 768-576 Desktop to a 768-576 TV :confused:
Maybe this is because the Geforce chip is giving the TV out chip a better source to resample.
I use level 3 flicker filtering in TVtool.
This combined with ffdshow filters gives me the most Film like viewing that I have seen over 3 years of tweaking, I have a sharp but natural image with no shimmering on highlights and almost no flicker.

Anyway I'm getting of topic.
My point is, if I can see a clear improvement using ffdshow resize with sharpen using my lowly display device, it must be of use to others.

Regards,

Owen
post #12 of 168
I had a crack at this and had some interesting results.

First I run 1440x960@60 for NTSC and either 1440x768@75 for PAL I can use 1440x960@50 for pal.

I resized to 1024x768 which looked grat on my NTSC res but it screwed my aspect ratio on pal res.

I suspect because it's a non standard aspect.

However my first impressions were it improved things.

CPU hit is high and stutter increased.....particularly on pal res it stuttered constantly though it could have been for reasons of aspect ratio.

Should I be resizing to the actual desktop res? ie 1440x960?

What settings do you use on luma sharpening ?

I also used Dscaler gradual denoise.

I have radeon 9700 MP1 and Barco 808 cpu is 2GHz p4, 768MB.
post #13 of 168
Thread Starter 
Rickd,
I suggest you resize to you desktop resolution.
Thats high for a 2G P4 CPU.
That should fix you aspect ratio.
Dscaler gradual denoise is to slow at that res.
Use ffdshow Gradual Denoise at about 30-40.
Denoise before you resize to save on CPU.
For sharpening I use 1.00 or even 2.00 for both Luma and Croma.
The sharpen effect is subtle, so you can use a lot.
If you can get this running I am shure you will be impressed.
If you are using VMR go back to Overlay mode to get back some CPU.
It is more important to get the resize and sharpen working then use VMR on a Radion. Your overlay is good enough.

Keep at it, its worth the effort

Regards,

Owen.
post #14 of 168
I thought the idea behind denoise was that it should be used AFTER sharpening to take care of any noise introduced through the process.
post #15 of 168
Thread Starter 
Nima,
Yes the denoise can be use after resize and sharpen but it dose not make much differance.
Resize filter dose not introduce noise but tends to reduce it and gives a very smooth yet detailed image.
The important thing is to resize as high as your CPU with allow and then Croma and Luma sharpen.
When we can buy 4 Gig CPU's we will have the option of denoising after resize.
If you have not done so try it.
Everyone interested in high quality should try it. And who isn't.

Regards,

Owen
post #16 of 168
You should resize to your Display Output resolution. If your projector supports 1440x960, use that. But don't oversize beyond the resolution so that scale-down must occur. FFDShow has an option to maintain aspect ratio on resize, that may work better for you.
post #17 of 168
So is my Radeon not scaling anymore then ?
And if so, is this software scaling superior to the Radeon's hardware scaling ? :confused:
post #18 of 168
Thread Starter 
Nima,
More action, less talk. :D
Please try using Resize with Croma and Luma Sharpen and see how you like it. It costs nothing to try.
I would like some feedback from Radion users.

Blight,
I tried the 30.82 drivers. But found them no improvement over what I was using (28.x I think).
I then tried new beta 43.0 drivers, just for fun.
Performance is no different, but color balance is much improved.
With the older drivers I needed to reduce Red levels and was never quite happy.
Very noticeable on my still Photo's of known color balance.
With 40.3 color is perfect at defaults. Good one Nvidia, best picture yet.
All tests under VMR9. I will never go back to Overlay on a Geforce.
No problems have been encountered.
I now believe that Geforce is now a match for Radion.
It’s hard to see how color could get any better.


On the Resize issue, I want to get people to TRY and resize above there normal desktop resolution. It sounds counter productive to resize up and have the video card rescale down, but try it anyway.
The further I upscale the better I like it.
Compression artifacts are reduced and Sharpening is more subtle.
This may only be true for my setup, but I want others to have a go. It’s free.
At a minimum, use your desktop resolution if it is higher than your source resolution.
You must upscale the source, then Luma-Croma sharpen.
I am up scaleing PAL 720-576 to 1280-1024 in ffdshow and then down scaleing to 720-576 in video card for output to TV and it works great.
So everybody, experiment with different resolutions and make up your own mind.

Regards,

Owen
post #19 of 168
G'Day :D

Well, actually I tried it. There was no need to sell it so hard ;).
The thing is, I much would like to have the things I asked clarified:

The software vs hardware scaling!
The what do all the different sharpen filters each on their own do thingy?!
What do especially chroma and luma sharpen do because they look better than unsharp mask on a first observation :p
And most off all, why the heck do we spent so much time in here LOL
post #20 of 168
Thread Starter 
Nima,

Quote:
The software vs hardware scaling!
Your the one with the Radion. You tell us. :D
One thing though. If you use the video card to resize, any sharpening filters will be before the resize. Thats bad.


Quote:
The what do all the different sharpen filters each on their own do thingy?!
Can't answer that. Blight or anyone ????

Quote:
What do especially chroma and luma sharpen do because they look better than unsharp mask on a first observation
Can't answer that one either. It would be nice to know. Milan etc HELP....

Quote:
And most off all, why the heck do we spent so much time in here LOL
LOL. If anyone can answer that, they get the prize. :D

Regards,

Owen
post #21 of 168
seriously this is almost embarrasing. I've been using asharp and I'm perfectly happy with it...yet i've been checking this thread everyday to see if there's something better. If only they started selling 1920x1080 DVDs :(
post #22 of 168
Ok, I did some experimentation and here's what I found. At first I tried to resize to 1440x960 figuring that a double of the native dvd would be a good thing, however, my cpu became seriously impaired (athon1800).
At 1024x768 my cpu could handle it, but only when i switched from bicubic resizing to experimental. I find that experimental is equal on quality (image as clear as bicubic), but it uses a few less cpu cycles than bicubic. What's weird though is that some of my avi's and mpgs had problems with experimental that they didn't have with bicubic (massive stutters).
I also tried 1440x480 and found this to actually look better than 1024x768 (a scaling issue I assume), however i needed to set the sharpening filters a little higher to achieve the same effect.
What's interesting though is that I always used the same output resolution to my tv (720x480p) yet still achieved excellent sharpening results through upsizing and letting my hardware scale it back down. I will try outputting at 1440x480p to my tv (it's all about the scaling) and see if that looks even better.
All in all I'd say that using resize with sharpening to be a better sharpening method than asharp or regular unsharp mask. I just wish it wasn't soooo hardware demanding!
post #23 of 168
Thread Starter 
llamameat,
Good to see you experimenting.
Everyone will have to experiment to find the best compromise between upscale size, sharpening and denoising filters to suit there display and CPU.
I have a secondary PC similar to yours. Athlon 1700XP with a GF4 440MX running in VMR9 mode that can handle Deinterlace, Levels, Gradual denoise and Resize to 1024-768 with Luma and Croma Sharpen in that order, and plays DVD's and full resolution Xvid perfectly smoothly.
Using Cinemaster 1.5 decoders and Reclock in Zoom Player.
That’s near maxed out for that system.
1440x960 is going to be tough.
I think pixel count is the limitation for CPU
Aspect ratio should not mater.
1024x768=786432 Pixels to process
1440x960=1382400 Pixels to process
That’s about 40% more Pixels. Ouch. :eek:
I'll try that res on my 3Gig P4 to test the limits of current CPU's .
But from my experience with 1280x1024 or 1310720 Pixels, on the 3 Gig P4, any noise filtering will be a hard ask in VMR mode.
With ffdshow Gradual denoise and Overlay mode, it may be possible.

What video card are you running?
Is it in VMR mode?

Regards,

Owen
post #24 of 168
Hardware vs. Software scaling:

There is no difference, it's all in the math, the hardware just does it faster. I'm not sure what interpolation the hardware does for image scaling. I think it's supposed to be Anisotropic for ATI Overlay (which is why I hate ATI), Bicubic for NVIDIA and Bicubic for all VMR9.

But heck, who really knows.
post #25 of 168
Owen, I have a geforce so I have to use VMR (overlay on geforce is downright awful by comparison).

Here's what I've discovered... The version of ffdshow marked as Dec 11 2003 takes considerably less cpu for resizing than any of the newer versions. I was using the jan03 version and my cpu was at 97% when I resized to 1440x480. Using the december version it went all the way down to 65% and the image quality looked exactly the same.

Now to try for 1440x960...buhaha
post #26 of 168
Thread Starter 
llamameat,

Good, you have confirmed my findings on ffdshow speed.
See my first post:

Quote:
NOTE: I find that versions of ffdshow after 12/13/02 are slow, so keep that in mind.
I have just tested 1440x960 Resize with Sharpen and ffdshow Gradual denoise before resize and it works smoothly on my system. Just. :eek:
Good luke.

Regards,

Owen.
post #27 of 168
Where do you get that version I only have seen the 03012003 version for dx 9.0.

All others seem earlier
post #28 of 168
Thread Starter 
Rickd,

You want ffdshow-20021213.exe

Available here:

http://sourceforge.net/project/showf...group_id=53761

Happy Viewing, ;)

Owen


PS: Please excuse the stupid US date format in my previous post :D
post #29 of 168
Thread Starter 
llamameat,

Since your a fellow Geforcer, Get your hands on the Detonator 43.0 beta drivers.
Color ballance is MUCH better in VMR9

What with ffdshow Resize with Sharpen, Gradual denoise, VMR9 and Detonator 43.0, I'm in video heaven. :D

Owen.
post #30 of 168
Blight,

>>There is no difference, it's all in the math, the hardware just does it faster. I'm not sure what interpolation the hardware does for image scaling. I think it's supposed to be Anisotropic for ATI Overlay (which is why I hate ATI), Bicubic for NVIDIA and Bicubic for all VMR9.

You're right -- as most of the hardware uses similiar scaling algorithms (bicubic scaling).

However, the advantage of a software approach is that you've got flexibility - you can use a different, superior, scaling algorithm. One, that might be able to do sharpening better than the hardware. You could even have adjustable transitions between DVD pixels - a more gradual blend or a sharper blend - because you're working with more pixels than 720x480 after you've software-scaled the video. That can have a good effect.
New Posts  All Forums:Forum Nav:
  Return Home
  Back to Forum: Home Theater Computers
This thread is locked  
AVS › AVS Forum › Video Components › Home Theater Computers › ffdshow Resize filter to Sharpen