Capture At or Resize Later?

[jjpslu]

Hi,

I've a few 2.5-3hr. iuvcr .avi caps at 704x480 48k NTSC and would like to obtain the BEST quality for each of these on its own DVD, (either compliant or "noncompliant"), for TV playback.

Which would be better during Encoding:
1. Encode at 704x480
2. Encode at a lower resolution - 352x480

I'm very close, but am missing that little bit of info that will make my end TV viewable product more acceptable than my previous results.

Please suggest your favorite guides/settings to accomplish this?

Feel free to email me directly with your assistance at jjp1@optonline.net.

Much appreciation!

[cosmin]

It's best to capture at the highest resolution - so 704x480 is fine. If your source is VHS, 704x480 is better than 720x480, because you avoid the vertical black bars on the margins, and the movie will be slightly easier to compress.

Next step: when encoding from a VHS source, the best bet is to encode to 352x480 (half-horizontal, full-vertical). Most MPEG encoders handle this conversion on the fly, so you shouldn't have to do it explicitly. Moreover, this resolution (named Half-D1) is accepted by the DVD standard.

Some people claim that encoding into 704x480 is better. However, this is true only for captures coming from TV broadcasts, or from a Super-VHS source. The horizontal analog resolution of the regular VHS is slightly below the digital correspondent given by 352 pixels. (For more info, search the net for "horizontal resolution of VHS"). You may, of course, encode a VHS capture into 704x480, but you have to spend extra disc space, and you gain nothing in return.

I hope this helps.

[NIIDeep]

Great.. this answered one of my questions...

So can I conclude the following for a VHS to DVD transfer?

- If I plan to do some extensive editing... I should cap at full D1, edit, and encode to MPEG2 at half D1.

- If I want to do a straight transfer I should cap to MPEG2 directly at half D1. Capping at full D1 to an MPEG2 file that's only going to be burned to a DVD later would give me no advantages. Is this correct?

Thanks for your help.

[bugster]

As well as what cosmin said, there is another thing to consider when choosing wether or not to use 1/2 D1. You said the caps are 2.5 to 3 hours. Thats quite a lot for a DVDr. So if you want all this on one disc you need to encode with a bitrate (Avg if VBR) af 3000 to 3500 Kbps. This is getting towards the low end for Full D1. Half D1 has half the pixels per frame, so even though the picture may end up a bit softer looking (due to the lower res), the chances of mpeg artifacts are much reduced.

[cosmin]

- If I plan to do some extensive editing... I should cap at full D1, edit, and encode to MPEG2 at half D1.

Yes.

- If I want to do a straight transfer I should cap to MPEG2 directly at half D1. Capping at full D1 to an MPEG2 file that's only going to be burned to a DVD later would give me no advantages.

Yes... almost.

Ideally, you should be able to capture at half-D1, encode to MPEG2, and obtain the results you expect. Unfortunately, it depends on the quality of your capture card. Apparently, with the mundane capture cards, you get better results if you capture at full resolution and resize later at half-resolution, than if you'd capture directly at half-resolution.

This may sound weird, but if you ask a digital photographer, s/he will tell you that you get a better image when you take a picture with a 4 megapixel camera that you resize later to 2 megapixels, than if you'd take the same picture with a 2 megapixel camera (assuming the two cameras have lenses of comparable quality).

Back to your question: if you have a high-quality capture card with a high-quality hardware MPEG2 encoder (such as Hauppage PVR), then you can capture directly at half-D1. Otherwise, if quality is a priority (assuming that you have enough hard-disc space), you'd better go with a full resolution capture, and with an offline, high-quality MPEG2 encoder (plenty too choose from, e.g. Canopus ProCoder, TMPGEnc, Cinemacraft Encoder or MainConcept) in which you scale the video from full-D1 to half-D1.

Last and not least, if you have access to a digital camcorder with analog input, you get far better results if you capture on tape, then you download the AVI via the FireWire card.

[NIIDeep]

Cosmin,

I am most grateful for your response. Thank you very much.

I have a ATI AIW 9800 Pro which has a h/w MPEG2 encoder. From what I understand it should be fairly high quality. I can understand your digital photography analogy to some extent since I dabble in that as well.

Thanks!

[trevlac]

Yes... almost.

Ideally, you should be able to capture at half-D1, encode to MPEG2, and obtain the results you expect. Unfortunately, it depends on the quality of your capture card. Apparently, with the mundane capture cards, you get better results if you capture at full resolution and resize later at half-resolution, than if you'd capture directly at half-resolution.

This may sound weird, but if you ask a digital photographer, s/he will tell you that you get a better image when you take a picture with a 4 megapixel camera that you resize later to 2 megapixels, than if you'd take the same picture with a 2 megapixel camera (assuming the two cameras have lenses of comparable quality).

This is reasonable, but not really correct.

It all has to do with how the digital image is process by any given device (including your PC). There are various ways to resize(resample) an image. Unfortunately, there is no rule of thumb on what does what. (Although we would like that ... i know. ) You have to know what your device does.

For a newer ATI card, 704x480 should be used if bitrate is not a problem. If it is, go to 352x480, knowing that you may be softening the picture (even VHS).

Bottom line: A source has a resolution which is probably different than the maximum resolution of it's format. As it goes thru the pipe to digital, back to your TV, each step can effect the quality. The only way to truely know is to test and see. Then you have to decide if the trade offs involved matter.

[jjpslu]

Any quick hints as to misc. settings or bitrates that might enable a sharper and quicker encode (from 704x480 to 352x480 2.5-3hr VHS cap)?

Thanx!
jjp1@optonline.net

[cosmin]

There are various ways to resize(resample) an image. Unfortunately, there is no rule of thumb on what does what. (Although we would like that ... i know. )

Then you should be happy to find out that there is a rule of thumb... in fact, there are two!

There is a theoretical rule of thumb, and a quite straightforward one: downsampling or upsampling by a factor N is done by dividing or multiplying the frequencies by N. The bad news is, this simple operation in the frequency domain translates, in the spatial domain, to a convolution with a function that is isomorphic to sin(x)/x, which has infinite support and requires lots of computation for the best discrete approximation...

Now in plain English An ideal (quality-wise) resizing is possible, but too expensive computationally. Compromises are necessary: you either want fast computation and you'll use a short filter that resembles the ideal one only slightly, and you'll end up with so-so results (at the extreme end, you get crappy results with a bland nearest-neighbor resizing); or you want good quality and you'll have to use a long filter to resemble the ideal one more closely - but which is slower to compute. Fortunately, there are filters that deliver a very good compromise: the computation is acceptable, and the quality is also very good. Bicubic interpolation (Mitchell-Netravali) and Lanczos are the most popular, and, perceptually, they are satisfactory for nearly all people. This also explains why (bi)linear interpolation is not that good: it's way better than nearest-neighbor method, but still not close enough to what's needed, and it softens your image.

So here is the second rule of thumb: the longer the filter, the better the quality. Of course, if you don't see the difference on the TV, it doesn't matter. I am just explaining why a good post-processing operation may make a difference when compared to a quick analog-to-digital conversion that downsamples on the fly.

As it goes thru the pipe to digital, back to your TV, each step can effect the quality. The only way to truely know is to test and see. Then you have to decide if the trade offs involved matter.

Correct.

[cosmin]

Any quick hints as to misc. settings or bitrates that might enable a sharper and quicker encode (from 704x480 to 352x480 2.5-3hr VHS cap)?

Bitrates depend on the quality of the source (a noisier source requires a higher bitrate), and also the final output quality that you desire. If you have a 352x480 source, you can try VBR 4000 kbps for a start. If you don't like that, you have to raise it. Given that the resolution is half, however, you might be just as happy with something even lower than 4000.

Or, if you want to fit it all on a single DVD, you should start from the other end. Use a bitrate calculator, figure out how much you can afford, and hope that the quality will be good enough for what you are limited to. Make sure you use a good MPEG2 encoder.

Best regards

[thecoalman]

If your source is VHS, 704x480 is better than 720x480, because you avoid the vertical black bars on the margins, and the movie will be slightly easier to compress.
.

Just wondering, I thought the full video, overscan and all was captured regardless of resolution selected.

[trevlac]

If your source is VHS, 704x480 is better than 720x480, because you avoid the vertical black bars on the margins, and the movie will be slightly easier to compress.
.

Just wondering, I thought the full video, overscan and all was captured regardless of resolution selected.

This is not the case. There is a big section in the doom9 capture guide that covers what happens on various cards. For ATI cards, I believe the scan width captured is around 704 aka 13.5 MHz * 52.15 micro seconds. As I understand, newer ATI drivers only allow caps at 704x. BTW, this is a combination of hardware and driver ... but there is a way to test.

[vhelp]

@ jjpslu

Regarding your question.. I would capture at the highest your card will allow
and then resize down. But only if you intend to do so.
.
I have my own OP's on Resing etc. Here's what I feel is recommended.

* captured: 704x480 - you can do the following:
... keep as 704, or
... if DVD projects, then add boarders to make 720 (8 x 8 pixels) or,
... resize down 352x480

* captured: 688x480 - you can do the following:
... resize to 704 by adding (8 x 8 pixels) or,
... resize to 720 by adding (16 x 16 pixels)

* captured: 640x480 - you can do the following:
... resize to 352x480

* captured: 352x480 - you can do the following:
... keep it at 352 x 480

What if your source is from VHS tapes ??

Well.., again, my OP is like this.
capture as high your capture card will allow (for me, that would be 720)
and resize to 352 (if your goal is 352 x 480) otherwise, I would keep it
at 720 x 480 and encode and author it to DVD as such.
.
There seems to be a missing fact or something going on here w/ respect
to VHS being your source. If you are capturing a VHS source and using
720 as your capture resolution. I would not resize down to 352, when
all your dvd player will do is resize it anyways, as your final result
will look slightely blurry. So, I would keep it at 720 and encode it
to DVD. This way, you won't loose resolution, and your source will
maintain sharpness when you view it by your dvd player/tv set.
.
But, most important.., finalize your projects to what sutes your eyes
since no two pair of eyes see the same thing(s)


@ trevlac,

Sorry for the OT here, but..

I visited over to doom9 and read up on the post (see below) few weeks
ago, and again over this weekend. But, because of my pc crashing and
my loosing a great deal of (pretty much everything, including a number
of files and things) and my e-mail ability (email is down) I was
wondering if you have that sampele source ( .AVI ) file that has
the "test pattern" that I can burn to a DVD-R disk. I had one that
I burned (long time ago) but lost it. And, since I have no email
ability for the time being, I can't get my old password from doom9
to make posts there

* Yet another Capture Test ...

I wanted to test this out on my Winfast TV2000 XP "Expert" card.

FWIW and Case you didn't know, this card has gone through THREE revisions.
I had to pick up a 2nd one (other one non-functional) and found that
the card I have now is REV C (though romar had it, REV A had problems, and
REV B was the good one, which I had ..till it broke )
.
With my new REV C board, I'd like to run some tests, using your
pattern. I'd like the .AVI if you have it, so that I can encode it to
an MPEG-2 file and burn to DVD disk.
.
If you have this pattern, and don't mind posting a link to it for D/L, I'd
appreciate it
.
I hope that I can get my Winfast to work under my Windows 98 Gold setup.
I seem to be having problems w/ it installing (my REV B card) and it gives
me computer hangs. Anyways.

Cheers all,
-vhelp

[trevlac]

@Cosmin,

Well again, you are on the right track .... BTW: I believe you know much more than many. I'd just like to tweak what you have said, not really refute it.

Now in plain English An ideal (quality-wise) resizing is possible, but too expensive computationally.

No it is not possible. A video frame has a limited sample width. A sinc filter is infinite (as you stated). To convolve a FIR with a signal, for practical purposes, the filter must be much shorter than the signal, because you can only produce Input - filter tap samples. IE 720x 100 tap filter == 620 output samples with no resize. To get the 50 on each edge, you need to fake it by using grey, or repeating the edge pixel.

Of course Infinite is not possible, but really long is not practical either.

Compromises are necessary: you either want fast computation and you'll use a short filter that resembles the ideal one only slightly, and you'll end up with so-so results (at the extreme end, you get crappy results with a bland nearest-neighbor resizing); or you want good quality and you'll have to use a long filter to resemble the ideal one more closely - but which is slower to compute. Fortunately, there are filters that deliver a very good compromise: the computation is acceptable, and the quality is also very good. Bicubic interpolation (Mitchell-Netravali) and Lanczos are the most popular, and, perceptually, they are satisfactory for nearly all people.

Frankly, I don't know enough about bicubic to say too much, but it is a polynomial equation that allows curve estimations. Bottom line, it is a way to estimate the position of new samples when doing interpolation.

Lanczos on the other hand does take the sinc function (which is used because it filters out all duplicate frequencies due to the sampling process. AKA .. pixels are not squares and we never see pixels even though most people think we do.) Lanczos is a truncated sinc (sin(x)/sin). Lanczos 3 truncates at 3 on each side of a given pixel (aka 6 taps) and Laczos 4 truncates at 4 (aka 8 taps). There is then a smoothing that is done to the trncated sinc. I believe some other methods of this smoothing (windowing) are blackman ... kaiser ... can't recall.

As a comparison, the BT878 chip which can be had on a sub $30 capture card uses a 32 tap interpolation filter when it resizes a capture. Another thing to know, capture cards always resize a capture. This is because they need to oversample for many reasons. So you always get a down sized version of what is originally sampled.

This also explains why (bi)linear interpolation is not that good: it's way better than nearest-neighbor method, but still not close enough to what's needed, and it softens your image.

Bilinear is a simple average between 2 pixels. AKA a straight line, no curve, a 1 tap filter. It is an easy method but as you say, poor. Nearest neighbor (point sample) can acctually be good. If you know there are no high frequencies in your picture, and you are going to decimate (downsize) by an interger amount, NN will give a result as good as a sinc. This result is better than Bicubic and Bilinear because the known samples are left alone in NN and changed in the others. In a Sinc, the known samples are not changed, only new ones are interpolated. You have to know (or assume) you have no high frequencies so you do not cause aliasing artifacts. If you do, you should smooth them out.

So here is the second rule of thumb: the longer the filter, the better the quality. Of course, if you don't see the difference on the TV, it doesn't matter. I am just explaining why a good post-processing operation may make a difference when compared to a quick analog-to-digital conversion that downsamples on the fly.

As I stated above, a cheap chip has a longer filter than any video post processing 32 vs 8. Actually, the best filters are non-linear. They don't use convolution. They use logic to determine the edges and keep the smooth parts smooth and the sharp parts sharp.

Bottom line ...
There is theory and practice... the therory is good to understand the general process. In the practice you find that there a quite a few unknowns ... like the frequencies in a given source ... like the type of filter used by a given piece of hardware ... like the true limits of using a filter on video ... like is anyone going to notice.


BTW: I have played with writting a variable tap convolution filter. Avisynth source is a good example of various resize filters. I can point you to this stuff if you care. Helped me clearify a few things.

Also, I wanted to say that people put too much thought into resolution. Good color (proper blacks, vibrant and true colors, etc ) makes much more of a difference after a certain point in the quest for the best resolution.

Cheers to you and I hope I don't sound too preachy ...

[trevlac]

But, most important.., finalize your projects to what sutes your eyes
since no two pair of eyes see the same thing(s)

Yes Yes ... you must really test and see. It is good to get some guidelines from others, but one really needs to then try it and make sure it works and they know what you get.


-V

http://trevlac.us/pics/index.html

There are a bunch of pics there. I'm not sure of the one you want. It's either the resoluiton1.png or the crop.jpg. I think.

[vhelp]

@ trevlac,

Thanks

yeah, at the moment, I'm head on out to the music store - getting Dave Mathew's
solo cd (something "Devil" or something like that) anyways.

One more dumb question regarding your tests..

Have you done a test sample w/ an SVHS machine yet ??
I'm curious to see the results of a VHS vs. SVHS. I think that would make
an interesting find
.
My units' the JVC S-VHS HR-S3910U. Anyways.

Also, FWIW.. I have ben testing my Winfast capture at 688 and resizing
to 720. I think they look pretty good. I've ben testing this resizing on
some of my Star Trek widescreen VHS tapes and w/ good results

Keep a watchful eye out for my response, if I ever get my Winfast card
working again.

Thanks again,
-vhelp

[vhelp]

..oh, before I leave, what's the AVIsynth script that takes a bitmpat (or your
.png pic) and produces a number of frames, so that I can frameserve it into
TMPG for encoding ?

-vhelp

[trevlac]

@jjpslu

Sorry for taking this off topic ... but hopefully people have given you the info you need. I'd say on the bitrate / sharpness thing, that these are opposing items. Smoothing is probably a good thing if you want 3hrs of vhs on a dvd.

@V

I provided Wilbert some vhs vs svhs vs digital sat shots for the doom9 capture guide. Look there. I have the same vcr as you.

The avisynth cmd is imageReader (i think) which is built in. Sorry, but you have to do the syntax yourself. You've been here a while... Remember Truman? He provided an AVS guide a long time ago. Probably an image reader type of example in that too.

Trev

PS: i've been here sooo long, i remember reading posts by adam where he sounded like a hot head complainer bitching at the mods. no kidding.

[jjpslu]

@Trevlac:

No problem!

I just seem to be missing "that little tweak" to finally get the finished product that I want. I really don't think that there is much anyone could offer at this point short of a few screenprints of what I am looking for.

Unfortunately, there isn't anyway of asking for such a thing while proving that I've actually spent hours each night attempting to perfect the process, but just lead that little bump forward. Sooo...balancing an all-inclusive post is a delicate feat.

Hopefully, someone here, someday will be in need of tweaking Vettes for massive, additional horsepower, that I can trade info with

[cosmin]

@trevlac -- is this Calvert spelled backwards?

Thank you very much for your additions. In particular, I didn't know that a $30 chip uses a 32-tap filter bank. I am pleasantly impressed.

I don't think we disagree on anything, after all. In my previous post, I over-simplified a couple of things (and I already feel we hijacked this thread too much...) so I didn't go into details about the relaxed requirement for a filter - that is, it needs not resemble the sinc() function, as long as its Fourier transform has near-compact support. This is why, for example, the Cohen-Daubechies-Feveau 9/7 wavelet and scaling function (used in JPEG-2000) is reportedly better than a 32-tap filter bank: this compactly-supported wavelet is nearly orthogonal.

If the source is high quality, then yes, there may be disagreements among picky people. I don't think this is the case when using a VHS source, though. Anyways, try and see. But as far as I know, arguing over good methods (e.g. bicubic vs. Lanczos) is like arguing over CCE vs. TMPGEnc, or Honda vs. Toyota

As for the edge detection issue, yes, of course you can employ high-level vision algorithms. I was referring just to classic DSP. Otherwise, you can go up to the extreme with the edges: encode the image using fractals, then resize it to whatever you want, and there will hardly be someone who'd think that there are other methods with better-looking results.

Cheers!

[trevlac]

@trevlac -- is this Calvert spelled backwards?

Why yes it is Here is another one for you ... nimsoc

I don't think we disagree on anything, after all. In my previous post, I over-simplified a couple of things (and I already feel we hijacked this thread too much...) so I didn't go into details about the relaxed requirement for a filter - ...

Gosh ... i can hardly dissagree on the wavelet thing ... I don't know the topic.

If the source is high quality, then yes, there may be disagreements among picky people. I don't think this is the case when using a VHS source, though. Anyways, try and see. But as far as I know, arguing over good methods (e.g. bicubic vs. Lanczos) is like arguing over CCE vs. TMPGEnc, or Honda vs. Toyota

Well, in the end ... it's always about the details. Personally, I think people should really learn more about proper color balance and such. Makes much more of a difference than resolution after a point.

As for the edge detection issue, yes, of course you can employ high-level vision algorithms. I was referring just to classic DSP. Otherwise, you can go up to the extreme with the edges: encode the image using fractals, then resize it to whatever you want, and there will hardly be someone who'd think that there are other methods with better-looking results.

Really ... i did not mean to refute you. Or even irritate you with my tone. I hardly have a handle on 'classic DSP' let alone using fractals. After all, i only have a simple BS degree ... and am a bit slow picking up these things. :P :P

But in the end ... past the theory ... some simple tests give really good answers.