704 vs 720, why crop?

[NamPla]

Yamakawa ... padded the 704 up to 720 while playing.

Liteon ... scaled the 704 up to 720

Both players treated the 352 pixel wide images the same way they treated the 704 pixel wide images.

I'm still trying to work out what your saying here...

(I know, it's been a long night)... heh... So is Ok to downsize from 720 to 352??

[junkmalle]

Yamakawa ... padded the 704 up to 720 while playing.

Liteon ... scaled the 704 up to 720

Both players treated the 352 pixel wide images the same way they treated the 704 pixel wide images.

I'm still trying to work out what your saying here...

(I know, it's been a long night)... heh... So is Ok to downsize from 720 to 352??

There probably is a "right" way but since we don't have the official DVD specs, DVD players handle the video differently, software players differ, TVs are usually out of whack, the differences in aspect ratio and image quality are minimal, and given that digital video is an inexact replication of the original analog world to start with (so there is no perfect reduction method), it probably doesn't matter much in most circumstances. You said earlier that you tried both and didn't see much of a difference. I'd go with that observation.

[iantri]

(Jumping into a conversation late..)

That document is extremely accurate, but complicated.

For more assistance wrapping your head around the issue, see this Doom9 guide, sections 3 though 5.

In summary, The video signal always has 525/625 lines, of which we use 480/576. The horizontal information is an analog wave; it is sampled to get however number of pixels digitally.

A 720 pixel image actually has the same pixel aspect ratio (or should) as a 704 pixel image, just with more picture information on left and right sides (well, most capture cards just pad it with black but nevermind that). This has nothing to do with overscan.

A 352 pixel image has half the image size horizontally of a 704 pixel image.

That is why going from 352 to 720 is slightly wrong; also why going from 720 to 352 is slightly wrong; this is why the AVIsynth manual demonstrates the cropping.

Problem: Most capture cards are poorly designed; they mis-sample and give either too much or too little image pixel per pixel. There is a complicated way of measuring in the doom9 guide.

Whether or not you should use 704x480/576 or 720x480/576 depends on which way is closest; for me with my BT878 with btwincap drivers I actually get the equivilent of 712 pixels at the proper aspect (just stretched out to 720), so I need to resample to 712 and pad 8 pixels. This is preferable to unnecessarily trimming 8 pixels to get down to 704.

This is different for each card; some do better than others. They really should be capturing more of the image at 720, but most just sample the same amount as for 704 for 720, this results in the same amount of picture, just stretched more. If your card does this, you SHOULD NOT be cropping anything to drop to 352. Though you should correct the aspect, since it is out slightly, if you are keeping it 720. The amount to correct depends on the card. Sucks, doesn't it?

Because of the differences between cards, people's results will differ.

I'd like to note that DV devices ALWAYS give the proper amount of image -- it is just crap consumer-level capture devices that screw it up.

It is REALLY REALLY complicated; but when it comes down to it unless your capture card is GROSSLY off (and most of then are only off 5-10 pixels), it doesn't really matter all that much.. your TV is likely worse calibrated than that.

Well, I guess that wasn't much of a summary.. if you aren't really sure still, don't lose any sleep over it.

[petar]

Alan69,

Don't take things personal. Nobody said you're wrong (at least I know I didn't), just that you didn't make a fair comparission. Resizing 704 to 352 will no doubt give better result. But we are talking resampling here.

Bicubic and Lancos3 are resizing filters. While they may even out the distortions and cause a more pleasing effect on many pictures, they are not 'more correct' or 'correcting' anything. They are adding an additional distortion, and it does not reduce things it only spreads out the other distortion. 352 cannot display the lower right circle, and it actually takes even more distortion to come up with the graying out vs the wider banding. It is coming up with it's gray by adding more blurriness. While it may seem more correct to some of you to end up with gray, note that this will add yet more softening on a normal picture, as well as adding more encoding time. It only seems more correct for that one case, over an entire random picture it will slightly worsen the picture.

Of course. Resampling algorithms distort the picture even more and they create a new one that has nothing to do with the old one. As an analogy, it's as if I were to give you a painting of 1m x 1m and tell you redraw the same thing but in 50cm x 50cm. Your new painting will probably be different in many aspects from the original but (if you're a good painter) will still be a good painting and someone that hasn't seen the original will find it pleasant to the eye.

On the other hand, if you were to mathematically resize a picture without resampling (i.e. without modifying its data), you'll get a more accurate image of the original but not as pleasant to the eye. You'll get more pixelation, stuff like stair stepping, jaggies etc. This is reduced significantly if you resize in a fixed ratio, but for video it's still better to resample rathen than resize. Takes longer, distorts the image but it gives a more "real" picture.


iantri,

A 720 pixel image actually has the same pixel aspect ratio (or should) as a 704 pixel image, just with more picture information on left and right sides (well, most capture cards just pad it with black but nevermind that). This has nothing to do with overscan.

A 352 pixel image has half the image size horizontally of a 704 pixel image.

That is why going from 352 to 720 is slightly wrong; also why going from 720 to 352 is slightly wrong

True for regular DVD-Video hardware players but not true for software players and hardware players capable of resizing. The only logical explanation, for people like me that will very unlikely spend money on the DVD specs, is that the DVD specs are flexible in this respect because there's only 2% difference, something that is probably irrelevant considering that most electronics isn't that accurate anyway.

The problem with some poor capture cards is not really connected to this. Assuming you have a DV video and you crop it to 704 (and then resize to 352 or not, doesn't matter), you'll get a different image on different players. I guess that's why all commercial DVDs stick to 720. But the difference is so small that I agree with you - we shouldn't lose any sleep over it .

[Wilbert]

On the other hand, if you were to mathematically resize a picture without resampling (i.e. without modifying its data),

??? Mathematically resizing is not modifying data?

[trevlac]

(BTW the Romans were shipping cocaine, marijuana, beer, etc... not just fish sauce! )

The fish sauce was probably just to confuse the dogs ...

[trevlac]

The only logical explanation, for people like me that will very unlikely spend money on the DVD specs, is that the DVD specs are flexible in this respect because there's only 2% difference, something that is probably irrelevant considering that most electronics isn't that accurate anyway.

Well, this is where we differ. Fair enough .... Another logical explaination is that Software DVD players are not properly/fully implementing the spec. DVD Demystified seems to imply this ... we may never know.

I did look into this a little, and appearently an mpeg stream can have a horizontal size and a horizontal display size (which is carried in an extension field). This all could be an mpeg thing, not specific to DVD specs. The iso docs are on neuron2's site. From a quick read ... it seems to 2 blocks on the sides of 720 may actually not be used. I asked for a tool that reads this info in a stream at the virtualdub board. A mod there 'fccHandler' recently posted that the color conversion matrix is actually carried in the stream (flag choise of about 4 types). This info is in the same spot as the display size stuff.

Unfortunately mpeg is much looser than the part the DVD uses. Now with DVB, it is hard to get clear info on what applies to what.

[petar]

On the other hand, if you were to mathematically resize a picture without resampling (i.e. without modifying its data),

??? Mathematically resizing is not modifying data?

Don't take things out of context. What Alan69 was refering to is resizing without resampling. If you resize 704 to 352, or course you're modifying the image - you have half the pixels to work with. But with resampling you're modifying each pixel's data.


trevlac,

Another logical explaination is that Software DVD players are not properly/fully implementing the spec.

True, not arguing with you there. Might as well be the case. But junkmalle just confirmed that one of his hardware players does the same thing. As you say, we may never know .

[trevlac]

trevlac,

Another logical explaination is that Software DVD players are not properly/fully implementing the spec.

True, not arguing with you there. Might as well be the case. But junkmalle just confirmed that one of his hardware players does the same thing. As you say, we may never know .

Ah ... but that Lighton is most likely a software player in disguise. But again ... you might be correct. Software players may be able to do more than the hw ones ... so they do ... both in spec. But I don't think so.

On the other hand, if you were to mathematically resize a picture without resampling (i.e. without modifying its data),

??? Mathematically resizing is not modifying data?

Don't take things out of context. What Alan69 was refering to is resizing without resampling. If you resize 704 to 352, or course you're modifying the image - you have half the pixels to work with. But with resampling you're modifying each pixel's data.

I slightly disagree with this ... although, I am no expert.

Summary - 1/2x sample pixels don't line up with 1x sample pixels. The 1/2 sample pixels 'cover' a wider area with a point measurement. NN does keep the old pixel, but this is not the measure we want. The other methods calculate the new pixel.

How I understand it ...

If the pixels are a representation of an analog wave (which for us they are) ... resampling is a 'better' method of a resize. Bicubic, Bilinear, Nearest Neighbor can all be implemented as a resample.

When you do NN and throw away 1/2 the pixels, you are introducing a spatial distortion.

__________|__________ A
   x    x    x    x   B
      o        o      C

A - The analog line with center identified
B - 2x samples
C - 1x samples

NN - Assumes the o's are the same as the x's. Although it looks like you get to keep x's you already had, you really wanted o's.

NN duplicates. Bilinear takes an average. Bicubic estimates the o. Lanczos calculates the o (but it fudges a bit so as to not have an unrealistic calculation).

----------------------
Down sampling is a streatch of my knowledge so be critical.

On the other hand, Up sampilng is quite clear to me ... and shows the flaws in NN. In up sampling ... Lanczos keeps the known samples unchanged. So the statement "But with resampling you're modifying each pixel's data." is not how I see it.

Of course it could be argued that you can not faithfully recreate the analog from samples.


edit
If you want to see what Avery says about his Vdub resize methdos, it is in the following text. Unfortunately when he talks, he tends to leave gaps that my knowledge does not fill in. So read at your own risk

http://virtualdub.everwicked.com/index.php?act=ST&f=7&t=2791&hl=lanczos&s=83eab8ca38eb...be776d1aaf2c9c

[petar]

I slightly disagree with this ... although, I am no expert.

I wouldn't call myself an expert either.

I agree with what you said and goes with what I mentioned before. Don't know why you say you disagree (though only slightly, as you say ).

Maybe I'm just crap at explaining things. I'll try again.

When resizing 704 to 352 and you simply throw away half the pixels, each new pixel in the new image is equal to one pixel in the old one, hence you haven't modified any data. The colours will remain exactly the same. But with resampling (with the exception of NN) you average the values of the surrounding pixels (4, for example for bi-linear) so you may get that every single pixel in the new image will have completely different pixel values to those in the original. That's what I meant by "with resampling you're modifying each pixel's data". But that's normally a good thing to avoid break up of the image features caused by simply throwing away pixels.

[Wilbert]

I thought you agreed with Trev?

When resizing 704 to 352 and you simply throw away half the pixels (...)

This is in general not true. That is only the case if you use NN, like Trev said.

I'm a bit confused about your post, because later you say something which seems to contradict the quote above

(with the exception of NN) you average the values of the surrounding pixels (4, for example for bi-linear)

[trevlac]

Don't know why you say you disagree
......
When resizing 704 to 352 and you simply throw away half the pixels, each new pixel in the new image is equal to one pixel in the old one, hence you haven't modified any data.

I would call this nearest neighbor downsizing. You get a spatial distortion. The left over pixels are not centered on the part of the line that they are supposed to measure.

The colours will remain exactly the same.

But they will not be in the right spot.

But with resampling (with the exception of NN) you average the values of the surrounding pixels (4, for example for bi-linear) so you may get that every single pixel in the new image will have completely different pixel values to those in the original.

For lanczos, you only avg the values when you are trying to calculate a new pixel. If the old pixels are in the right spot, you keep them. 720->325 may acutally keep some pixels for lanczos. 704->352 does not because we are always talking about samples in a different spot.

That's what I meant by "with resampling you're modifying each pixel's data". But that's normally a good thing to avoid break up of the image features caused by simply throwing away pixels.

This with bi-linear is why I only partially disagree. Bilinear does avg pixels. But that is probably more proper than NN.

Maybe I'm just crap at explaining things. I'll try again.

I think you are pretty good. Hopefully others understand.

[petar]

I thought you agreed with Trev?

When resizing 704 to 352 and you simply throw away half the pixels (...)

This is in general not true. That is only the case if you use NN, like Trev said.

I'm a bit confused about your post, because later you say something which seems to contradict the quote above

(with the exception of NN) you average the values of the surrounding pixels (4, for example for bi-linear)

Again you are quoting things out of context and twisting my words.

OK, I'll explain myself as simple as possible and please stop chopping half-sentences because this is getting ridiculous. If you're knowledgable on this topic, I'm sure you understand what I mean.

My whole point was that resizing by throwing away pixels (which is NN) doesn't change the pixel values, and though is fast and good to preserve hard edges (for computer generated images for example) is not a good method to use to compare downsized "real" videos. Resampling (with the exception of NN which does the former) assigns new color values to any new pixels it creates, based on the color values of existing pixels in the original. Hence the pixel data changes. Doesn't mean it will happen on the whole image (it may) but that's the idea behind it. Still it's better for "real" videos.

So, I was just trying to distingush between the two. I may not have used the best terminology, but I'm sure you can get the idea behind it.

Alan69, when testing the resizing of 720 to 352 and 704 to 352 used the former. I was just trying to point out that though it will give much better result for 704 to 352 it is not a fair comparisson because the resizing method is not very good for video.

[Wilbert]

Sorry, I see that I misread the posts of you. I thought I saw the following

When resizing 704 to 352, you simply throw away half the pixels (...)

Note the difference. But, that's indeed not what you said. My apologies for not reading properly.

[petar]

It's cool . English is not my first language so I'm sure I'm not always clear.