New deinterlacing King
- Status
1 - 20 of 72 Posts
Joined
·
1,835 Posts
SpongeBob SquarePants?
Joined
·
2,006 Posts
Jeff: I guess there is a new dscaler filter...as long as it isn't sideshow BOB we should be OK. Having said that you know where you are with sideshow BOB. With my HTPC it performance is a lot less reliable!
Gordon
P.S. Really must pop in to see your system!
Gordon
P.S. Really must pop in to see your system!
Gordon,
Nothing gets past you eh? Yes, this is John Adcock's new adapted version of TomsMoComp and very nice it is too. I guess John having DCDi to compare to is starting to pay off because from what I have seen so far this is a real cracker.
You are always welcome to come and have a look, but I know its a long long way.
Jeff
Nothing gets past you eh? Yes, this is John Adcock's new adapted version of TomsMoComp and very nice it is too. I guess John having DCDi to compare to is starting to pay off because from what I have seen so far this is a real cracker.
You are always welcome to come and have a look, but I know its a long long way.
Jeff
Quote:
Way "back in the day" DScaler had a Simple Bob deinterlacing method. The name evoked images and sometimes jokes on the DScaler mailing list. So when I wrote TomsMoComp I added a somewhat more elaborate method for 'bobbing' when there was fast (non-compensatible) motion and called it Wierd Bob.
But even that one could maybe be improved somewhat.
So JohnAd made an optional replacement for it and naturally called it Strange Bob.
In the spirit of things about a year ago when I was writing Greedy/HM I first developed the code for the vertical filter separately and named it Robert. But it was never released separately.
I guess it's becoming a DScaler tradition now.
- Tom
| Is Bob (the person) the new deinterlacing king or Bob (the bob/weave technique of deinterlacing)?! |
But even that one could maybe be improved somewhat.
So JohnAd made an optional replacement for it and naturally called it Strange Bob.
In the spirit of things about a year ago when I was writing Greedy/HM I first developed the code for the vertical filter separately and named it Robert. But it was never released separately.
I guess it's becoming a DScaler tradition now.
- Tom
Jeff
Thanks! It is still not really finished yet and I'm sure the parameters could do with some tweaking but overall I think we are at the very least closing the gap with DCDI
Li On
Strange Bob requires SSE support does your processor support SSE? otherwise it should be a fairly noticable difference on video sourced stuff?
Anybody called Robert
Sorry....
John
Thanks! It is still not really finished yet and I'm sure the parameters could do with some tweaking but overall I think we are at the very least closing the gap with DCDI
Li On
Strange Bob requires SSE support does your processor support SSE? otherwise it should be a fairly noticable difference on video sourced stuff?
Anybody called Robert
Sorry....
John
It may amuse some of you to know that I haven't actually used Strange Bob much yet as my current HTPC doesn't run it properly along with film detect without dropped frames. I did the coding on a machine without a capture card so I've only really seen it with stills
Anyway I'll be upgrading to a 2.4GHz P4 over the weekend so I'll finally be able to run it properly and hopefully enjoy some telly instead of poking around with MMX assembler which is something which only the most masochistic developers (me and Tom) do and which should be a clear signal that I need to get out more.....
John
Anyway I'll be upgrading to a 2.4GHz P4 over the weekend so I'll finally be able to run it properly and hopefully enjoy some telly instead of poking around with MMX assembler which is something which only the most masochistic developers (me and Tom) do and which should be a clear signal that I need to get out more.....
John
That sound nice but can it process on a diagonal path like DCDi?
DCDi or Directional Correlational Deinterlacing, eliminates the jagged edges that appear when standard interlaced video is viewed on progressive scan displays. During the initial launch of the HDTV industry, broadcasters recognized that a means of very high quality upconversion of NTSC was required. Faroudja entered this market immediately, developing the DCDi algorithm to eliminate the staircasing effect that conventional upconverters introduced to diagonal edges.
Artifact's were and are eliminated by angle correlated vector interpolation algorithms, developed through multiple and laborious experiments. Typical motion video processing relies on intra-field spatial interpolation to generate the missing lines of video from interlaced sourced material because it suppresses motion artifacts. However using spatial interpolation on moving images and diagonal edges (actually, all edges which are neither horizontal or vertical) creates a jagged edge artifact, since the missing pixels are interpolated from those lines above and below, which do not correspond to the edge itself.
Faroudja’s DCDi algorithm identifies all the moving edges in a scene and adjusts the angle of interpolation at each pixel so that the interpolation always follows the edge instead of crossing it, eliminating staircasing or jagged edge artifacts. The result is a smooth and natural looking image.
Can this BOB de-interlacing adjust the angle of interpolation?
DCDi or Directional Correlational Deinterlacing, eliminates the jagged edges that appear when standard interlaced video is viewed on progressive scan displays. During the initial launch of the HDTV industry, broadcasters recognized that a means of very high quality upconversion of NTSC was required. Faroudja entered this market immediately, developing the DCDi algorithm to eliminate the staircasing effect that conventional upconverters introduced to diagonal edges.
Artifact's were and are eliminated by angle correlated vector interpolation algorithms, developed through multiple and laborious experiments. Typical motion video processing relies on intra-field spatial interpolation to generate the missing lines of video from interlaced sourced material because it suppresses motion artifacts. However using spatial interpolation on moving images and diagonal edges (actually, all edges which are neither horizontal or vertical) creates a jagged edge artifact, since the missing pixels are interpolated from those lines above and below, which do not correspond to the edge itself.
Faroudja’s DCDi algorithm identifies all the moving edges in a scene and adjusts the angle of interpolation at each pixel so that the interpolation always follows the edge instead of crossing it, eliminating staircasing or jagged edge artifacts. The result is a smooth and natural looking image.
Can this BOB de-interlacing adjust the angle of interpolation?
HD-DVDMAN
Yes, both TomMoComp normal mode (Weird Bob) and the newer Strange Bob mode do diagonal searches. We just haven't got a marketing department to write speil for us
John
How about this :
TomsMoComp eliminates the jagged edges that appear when standard interlaced video is viewed on progressive scan displays. During the initial launch of the HDTV industry, broadcasters recognized that a means of very high quality upconversion of NTSC was required. DScaler entered this market late, developing a couple algorithms to try and eliminate the staircasing effect that conventional upconverters introduced to diagonal edges.
Artifact's can sometimes be eliminated by using these new interpolation algorithms, both alogorthms were developed by amateurs working in their own time. Typical motion video processing relies on intra-field spatial interpolation to generate the missing lines of video from interlaced sourced material because it suppresses motion artifacts. However using spatial interpolation on moving images and diagonal edges (actually, all edges which are neither horizontal or vertical) creates a jagged edge artifact, since the missing pixels are interpolated from those lines above and below, which do not correspond to the edge itself.
DScaler’s TomsMoComp algorithms identify all the moving edges in a scene and adjust the angle of interpolation at each pixel so that the interpolation always follows the edge instead of crossing it, eliminating staircasing or jagged edge artifacts. The result is a smooth and natural looking image.
Yes, both TomMoComp normal mode (Weird Bob) and the newer Strange Bob mode do diagonal searches. We just haven't got a marketing department to write speil for us
John
How about this :
TomsMoComp eliminates the jagged edges that appear when standard interlaced video is viewed on progressive scan displays. During the initial launch of the HDTV industry, broadcasters recognized that a means of very high quality upconversion of NTSC was required. DScaler entered this market late, developing a couple algorithms to try and eliminate the staircasing effect that conventional upconverters introduced to diagonal edges.
Artifact's can sometimes be eliminated by using these new interpolation algorithms, both alogorthms were developed by amateurs working in their own time. Typical motion video processing relies on intra-field spatial interpolation to generate the missing lines of video from interlaced sourced material because it suppresses motion artifacts. However using spatial interpolation on moving images and diagonal edges (actually, all edges which are neither horizontal or vertical) creates a jagged edge artifact, since the missing pixels are interpolated from those lines above and below, which do not correspond to the edge itself.
DScaler’s TomsMoComp algorithms identify all the moving edges in a scene and adjust the angle of interpolation at each pixel so that the interpolation always follows the edge instead of crossing it, eliminating staircasing or jagged edge artifacts. The result is a smooth and natural looking image.
HD-DVDMAN
Ask away if his word will sway you, Mark has never really been up close to the video methods but I'm sure he'll have an opinion.
I still think it's too early to declare a new "king" of motion adaptive deinterlacing. Teranex and Snell & Wilcox can duke it out at the top end with true motion compensating stuff but at the lower end it's getting too close to call.
John
Ask away if his word will sway you, Mark has never really been up close to the video methods but I'm sure he'll have an opinion.
I still think it's too early to declare a new "king" of motion adaptive deinterlacing. Teranex and Snell & Wilcox can duke it out at the top end with true motion compensating stuff but at the lower end it's getting too close to call.
John
Quote:
HD-DVDMAN -
That is true as far as it goes. However TomsMoComp will only bother to do that in the event of uncompensated motion. To the extent a motion search turns up the reliable value for the missing pixels then we do not have to interpolate them from the current field.
WierdBob use an ELA (edge based line averaging) technique first to interpolate pixels along a constant line, which may also be one of 5 different diagonals. But this value will not be used if the motion compensation search is successful since that algorithm can provide higher detail.
This difference will be especially obvious on still scenes.
But by all means, ask whoever you wish, or go read the code in Sourceforge.
- Tom
| DScaler’s TomsMoComp algorithms identify all the moving edges in a scene and adjust the angle of interpolation at each pixel so that the interpolation always follows the edge instead of crossing it, eliminating staircasing or jagged edge artifacts. The result is a smooth and natural looking image. |
That is true as far as it goes. However TomsMoComp will only bother to do that in the event of uncompensated motion. To the extent a motion search turns up the reliable value for the missing pixels then we do not have to interpolate them from the current field.
WierdBob use an ELA (edge based line averaging) technique first to interpolate pixels along a constant line, which may also be one of 5 different diagonals. But this value will not be used if the motion compensation search is successful since that algorithm can provide higher detail.
This difference will be especially obvious on still scenes.
But by all means, ask whoever you wish, or go read the code in Sourceforge.
- Tom
Tom
Sorry the quote was a joke paraphasing from the original DCDi press release, it was not meant to be technically accurate.
Without spin, Strange bob does have rudimentary edge detection which certainly helps when selecting which diagonal is the best.
More seriously Tom I'm still not sure that increasing the search effort results in a better picture. Could you outline how the mocomp bit works again, it may be that Strange bob leaves the registers in an incorrect state allowing "better" weaves to come in that actually degrade the image. Without doing full object motion estimation I don't see how you can be bringing in extra detail as detail will look like an artefact when doing the tests to see if the new choice is better than the previous ones.
What I see happening with an increased search effort is that sometimes a pixel is brought in on a bad diagonal from the weave options.
John
Sorry the quote was a joke paraphasing from the original DCDi press release, it was not meant to be technically accurate.
Without spin, Strange bob does have rudimentary edge detection which certainly helps when selecting which diagonal is the best.
More seriously Tom I'm still not sure that increasing the search effort results in a better picture. Could you outline how the mocomp bit works again, it may be that Strange bob leaves the registers in an incorrect state allowing "better" weaves to come in that actually degrade the image. Without doing full object motion estimation I don't see how you can be bringing in extra detail as detail will look like an artefact when doing the tests to see if the new choice is better than the previous ones.
What I see happening with an increased search effort is that sometimes a pixel is brought in on a bad diagonal from the weave options.
John
Quote:
John -
I was well aware of that. I briefly considered trying to put on a PR-man's hat and reply with an explanation somewhat less nerdly. But it was just against my nature.
Quote:
Well the best example is a SearchEffort=1 where we only search 1 location. At SE=1 this is just an adaptive deinterlace (+whateverBOB). If the pixel value (of 0 motion) is unchanged from the previous field to the succeeding field then we figure it is pretty credible if it wouldn't also generate obvious comb artifacts.
Now take (imaginary) SE=2. Here we could check that 1 pixel to the left in the preceding field matched the value 1 pixel to the right in the succeeding field. If so then that value has a good chance of being in the middle on the middle (current) field.
As you increase SE you have a better chance of finding the actual motion but also a better chance of false hits, mistakenly guessing you know what is going on. My current view is that most anything above about SE of 5-9 or so is diminishing returns.
This is compounded by a possible bug that I've never been able to find. Above SE=5 we start also searching vertically and that seems to cause more stray hits.
(SE=6,7,8 all really go to 9)
Note vertical searches are less dense because of interlace so we are really 2 lines away on the search.
Anyway, folks on slower machines that run only SE=3 will probably get most of the benefit, just searching 1 pixel to the left, right, and center. But on my 2.4 Mhz P4 I still like it at 11, for no particular reason.
BTW, increasing SE to the point where you get dropped frames will make it seem MUCH worse, since the fields don't match. But everybody should use a value of at least 1 or 3 unless it is a very slow machine.
- Tom
| Sorry the quote was a joke paraphasing from the original DCDi press release, it was not meant to be technically accurate. |
I was well aware of that.
I briefly considered trying to put on a PR-man's hat and reply with an explanation somewhat less nerdly. But it was just against my nature.Quote:
| More seriously Tom I'm still not sure that increasing the search effort results in a better picture. Could you outline how the mocomp bit works again ... |
Now take (imaginary) SE=2. Here we could check that 1 pixel to the left in the preceding field matched the value 1 pixel to the right in the succeeding field. If so then that value has a good chance of being in the middle on the middle (current) field.
As you increase SE you have a better chance of finding the actual motion but also a better chance of false hits, mistakenly guessing you know what is going on. My current view is that most anything above about SE of 5-9 or so is diminishing returns.
This is compounded by a possible bug that I've never been able to find. Above SE=5 we start also searching vertically and that seems to cause more stray hits.
(SE=6,7,8 all really go to 9)
Note vertical searches are less dense because of interlace so we are really 2 lines away on the search.
Anyway, folks on slower machines that run only SE=3 will probably get most of the benefit, just searching 1 pixel to the left, right, and center. But on my 2.4 Mhz P4 I still like it at 11, for no particular reason.
BTW, increasing SE to the point where you get dropped frames will make it seem MUCH worse, since the fields don't match. But everybody should use a value of at least 1 or 3 unless it is a very slow machine.
- Tom
1 - 20 of 72 Posts
- Status
Join the discussion
Join now to ask and comment!