Help with real time IVTC (same concept as capture).

[pbiancardi]

I know this is a capture forum but I figure if anyone will know this answer you guys will.

My goal is to do real time Laserdisc viewing (not capture) using MadVR renderer as a really high quality upscaling solution to take 480i video to 2160p.

I go LD Player -> ADV 7842 eval board (for comb filter) then output 480i signal over HDMI to Elgato HD60 Pro and use PotPlayer or MPC-BE to watch the uncompressed preview stream with MadVR doing the real time upscaling.

The issue I am having is I cannot no matter what I try get proper deinterlacing / IVTC for the 480i signal so I have to first pass it through a Denon video processor to convert to 480p then into the capture then all is fine (except the massive GPU usage for 59.94 fps video).

I have even tried to send a 1080i HDDVD signal through the same capture chain and it will not properly IVTC.

I examined a DVD Rip that can be played and IVTC'ed properly in GraphStudioNext and it goes File -> Lav Video Decoder -> Renderer so I assume that is a good way to do it.

I used GraphStudioNext to build my own graph HD60 Pro Capture -> Lav Video Decoder -> MadVR and if I press play in GraphStudio I do get a picture so I assume it is working.

How do I take that Graph and force PotPlayer or MPC to use that chain? Right now in MPC all of my ripped files will use Lav Video as I set it up to do but the capture stream uses Smart Tee Video (which I tried to delete in Windows to force the use of Lav Video but that broke the streaming).

Any other suggestions for getting proper IVTC done in real time?

Not sure if the HD60 Pro is the issue (it doesn't actually say it supports 480i) so I ordered and will try an ATI 750 PCI to take composite direct from LD player, not sure that will make a difference?

[lordsmurf]

Realtime IVTC can't be done.
It will guess, not know, and make mistakes.

[pbiancardi]

Realtime IVTC can't be done.
It will guess, not know, and make mistakes.

You mean specifically for low res LD? Obviously DVD / HDDVD can happen in real time?

[Cornucopia]

Depends on what you mean by realtime. If you mean with near-zero latency - NO, it is NOT possible.
It might be possible if you have a buffer that captures mutliple frames-worth and compares the frames and re-organizes the frames if necessary, but then the latency to do this for normal framerate video would be in the range of 6 frames, or up to 250msec. And that is if it is with theoretical maximum efficency & speed. The nearest this has been done in "realtime" in hardware equipment is with pulldown flags in DVD & Bluray players, but those are not doing analysis of the frames' actual signal but rather reorganizing following the flags dictates (which takes near to no processing power). And even then, it doesn't show up except in the higher-end machines (where it can output true 24p/23.976p).

Scott

[pbiancardi]

Thanks I appreciate the info.

I do have a processor (Denon DVP-602ci) that does perfect 480i60 to 1080p24 but that's probably doing blending or other hocus pocus not true IVTC?

[Cornucopia]

The manual doesn't say. It IS guessing: that's what "cadence detect" is referring to. And it does have a buffer: that's why it has adjustable audio delay, so it can maintain sync due to the need for (variable?) buffering in the video side. Thus, it does also have additional latency which you probably arent aware of.
If the cadence is rock solid & completely consistent 2:3 pulldown, it could very well give you clear progressive IVTC output - for some segments. Anything other than that (the majority of material, in my experience), and it will either have cadence match reset glitches or it will fudge over those changes using field/frame blending (Your "hocus pocus"). And if it is minor or only during certain transitions, you may not be primed enough to be aware of it when it passes by in "realtime", but it is there, and once it is it will be a LOT harder to get out.

Best practices says: don't attempt this in realtime, nor the scaling, as the quality of conversion is nowhere near what can be done after in software, as many of these threads describe. Cap at 480i, then use appropriate tools to analyze the material and get you the best conversion.


Scott

[jagabo]

It's possible to open a capture graph with AviSynth so it may be possible to IVTC in real time (meaning while capturing, not zero latency) while capturing with it.

[lordsmurf]

Realtime IVTC can't be done.
It will guess, not know, and make mistakes.

You mean specifically for low res LD? Obviously DVD / HDDVD can happen in real time?

As mentioned, those are flagged, not the same.

It's possible to open a capture graph with AviSynth so it may be possible to IVTC in real time (meaning while capturing, not zero latency) while capturing with it.

ATI AIW could IVTC at capture 20 years ago, but it wasn't precise. It made mistakes. Nothing here has changed, as nothing with capturing has really changed in 15+ years. The same mistakes were not made on analysis passes pre-encode (TMPGEnc back then),

[pbiancardi]

The manual doesn't say. It IS guessing: that's what "cadence detect" is referring to. And it does have a buffer: that's why it has adjustable audio delay, so it can maintain sync due to the need for (variable?) buffering in the video side. Thus, it does also have additional latency which you probably arent aware of.
If the cadence is rock solid & completely consistent 2:3 pulldown, it could very well give you clear progressive IVTC output - for some segments. Anything other than that (the majority of material, in my experience), and it will either have cadence match reset glitches or it will fudge over those changes using field/frame blending (Your "hocus pocus"). And if it is minor or only during certain transitions, you may not be primed enough to be aware of it when it passes by in "realtime", but it is there, and once it is it will be a LOT harder to get out.

Best practices says: don't attempt this in realtime, nor the scaling, as the quality of conversion is nowhere near what can be done after in software, as many of these threads describe. Cap at 480i, then use appropriate tools to analyze the material and get you the best conversion.


Scott

I really appreciate the info.

Real time scaling does work well though, MadVR can do that with very minimal dropped frames which seem to be caused by the real time capture viewing. Maybe you would know why it struggles with real time when playing from ripped file is a breeze?

I'm not going to capture thousands of LDs so real time is the only way for me and this does work just have to settle for no IVTC I guess.

[Cornucopia]

"Realtime" = constraint in timing of input + constraint in timing of output = extra constraint in timing of processing
File-based = only constraint in timing of output, input to memory for processing can be VERY fast (esp. w SSD), processing time has some leeway.


Scott

[pbiancardi]

"Realtime" = constraint in timing of input + constraint in timing of output = extra constraint in timing of processing
File-based = only constraint in timing of output, input to memory for processing can be VERY fast (esp. w SSD), processing time has some leeway.


Scott

So forgetting IVTC entirely and just using standard deinterlacing what is the bottleneck for real time viewing of a capture card and is there anything that can be done to improve it? My CPU / GPU don't seem to be working that hard but with a ripped file I get all of MadVR's queues to fill up and zero dropped frames where with same type of file now streaming over capture card the queues don't fill and I get minimal dropped frames. Don't get me wrong it does work and is really high quality but of course I would like the dropped frames to be zero whether they are noticeable or not.

[pbiancardi]

Realtime IVTC can't be done.
It will guess, not know, and make mistakes.

I have been looking at Osprey 260e which appears to have IVTC built into the driver, that would not work either?

[Cornucopia]

1) I thought you said you weren't going to be doing captures? (that's what the osprey does - I have a bunch of 210s). Viewing is a feature, but it isn't the core object of application, so it is quite possible that frames might be dropped in the viewer portion.
2) It still uses a buffer (2-frame or 3-frame, so only applicable to "simple" cadence material), and it still "guesses", with accompanying mistakes. The documentation even says that one should not attempt ITVC unless sure of the 24p origin of the material.
3) that real-time vs. file-based comparison applies to MANY algorithms, applications and devices. It isn't specific to IVTC.

Scott

[lordsmurf]

Osprey is no more accurate than ATI AIW.
It guesses.
The buffers will just barf out whatever it can before choking.

Some tasks cannot be done realtime, and potentially never will. GPU could probably do it, but no GPU-based analog capture card exists.