Technique to decrease compression ratio in DV footage at DVinfo.net

Wayne Morellini · July 29th, 2005, 02:14 AM

You wouldn't believe this, I have figured out a way that should lower the compression ratio of Mini DV footage, but I don't know if the camera codecs will allow it.

When you use the cropped wide screen feature, it records black bars top and bottom. But black bars should take up very little compression space in any descent compression routine, leaving more for the rest.

I did the calculations weeks ago, but for PAL progressive (that should also require less compression space than interlace) you should get significant reduction in the ratio of 25%. At these ratio's a little reduction should matter a lot. Now if we look at Wide screen cinema formats 2.35:1, is around 43% reduction, that puts you into Pro DV ratio territory. You should be able to make up an lens hood to crop it.

The compromise is that for cinema projection you loss the extra resolution of a true 16:9 progressive mode, but retain a square pixel. But one of the big problems with DV compression is rapid complex movement (though nothing compared to HDV 1080i) so it might be possible to use cropped 16:9 for special effects (explosions etc) and true 16:9 for everything else. So cropped wide screen is not a complete loss after all.

So what do you think?

Follow this link on pixel shifting single chip:
http://www.dvinfo.net/conf/showthrea...561#post339561

Keith Wakeham · July 29th, 2005, 04:12 AM

I was of the understanding that mini dv was a constant bitrate dct compression and that under constant bitrate with dct each block has so many values like a pixel would but because of dct the black area would be given the same x of y values so it wouldn't spare up any extra data. The interframe compression like mpeg 4 saves space because it detects no changes between keyframes and doesn't write much data then, while intraframe still needs to write all the black data.

I'm no compression expert so I could be way off but wouldn't having black bars for such a wide format result in sevre loss of resolution. You'd be tooping out at like 720 x 200 then with 4:1:1. That is just an estimated pixels number, to lazy to calc the real one for 2.35:1

Wayne Morellini · July 29th, 2005, 08:21 AM

I think it was closer to 240 pixels. That's the downside, but then again it is the same resolution per inch as the horizontal measurement. The way they get the horizontal up is through resolution upscaling, so it is a compromise. We're not talking about HD here, we are talking about basic results and improving detail (unless you have uncompressed component instead) .

About the compression data rate being divided up into blocks, that is the sort of thing I was worried about, but if it is really good it should be intelligent enough to identify that all the surrounding blocks are the same, and compress it as an area, and re-distribute the bandwidth to areas with details. If anybody could give links to positively confirm that it works like you said, and does not have area based compression, that would be good.

I can't remember (it has been since the 80's/early 90's since I have read about compression) but I thought they used a couple of different methods in DV like compression.

Matthew Wauhkonen · July 29th, 2005, 08:36 AM

DV frames are indeed self-contained. DV has no problem with explosions, fast movement, etc. beyond motion blur, which would be a problem with uncompressed footage as well. Turn your shutter speed up to 1/1000th of a second and shoot progressive and you'll notice that everything is super sharp in each frame even during fast motion but your footage looks choppy. So that's just completely not an issue with DV. Never has been, never will be. HDV, however, does have a GOP-based codec (MPEG2) so it does struggle with fast motion (and, yes, I've noticed this in action and it can get very bad, although I still think the FX1 is a really cool product).

As for the "black bars" there has been a lot of argument over how smart the DV codec is. Since the hardware DV codec is different in different cameras, some have even claimed that some "dumb" cameras devote as much space to the black bars as to the rest of the frame, while some "smarter" cameras devote more to the rest of the frame. I don't buy it. Although there's no proof either way, I think that the entire frame is compressed so that each pixel is given the same amount of attention. This is a real time codec invented like 15-20 years ago. High quality GOP codecs (unlike HDV) take a long time to compress footage on very fast computers. DV compresses on real time in 200 dollar camcorders. Sorry, it's just not that smart. Furthermore, due to the 4:1:1 colorspace, there will always be moer "color" compression artifacts in a letterboxed image than an anamorphic one. I have a dvx100a and always shoot squeeze mode for 16:9 and can say, although only anecdotally, that it looks much better than letterboxing.

Wayne Morellini · July 29th, 2005, 09:04 AM

Comparing HDV codec compression to DV codec compression on a fast computer would give better results. Computer computational performance is a bit misleading, they are very inefficient compared to custom silicon used in cameras like the HC1.

If it devotes the same amount to each pixel, that is horrible and the picture detail should suffer badly. What I am saying actually should be relatively little overhead (compared to Mpeg2). So I think it is possible, but only specs on the inner workings of DV would tell.

I have read positive reports of problems with DV in complex motion/detail. On a limited set bandwidth codec, it should only be able to keep up with limited amounts of detail, especially if space is alloted tied to basic blocks. So the detail would be wrong, or worse.

Keith Wakeham · July 29th, 2005, 09:38 AM

It would be nice if we could find a way to test this theory. I don't know how though. would be interesting to see anyway. But for hdv it might work because of the GOP interframe structure rather then the intraframe. Problem with that is how do you get a 16:9 camera to shoor 2.35:1.

Wayne Morellini · July 30th, 2005, 07:13 AM

I think, unless somebody comes by and provides the evidence I will have to look it up on the web and find an accurate description. Unfortunately I don't want to put everything else aside for four hours to do it. Maybe tomorrow, or Monday. Only a usefull tip for others.

Wayne Morellini · August 4th, 2005, 06:41 PM

Quote:

Originally Posted by Keith Wakeham

It would be nice if we could find a way to test this theory. I don't know how though. would be interesting to see anyway. But for hdv it might work because of the GOP interframe structure rather then the intraframe. Problem with that is how do you get a 16:9 camera to shoor 2.35:1.

OK

Sorry for the delay been a bad week, and I'm not too well, as well. Also sorry for the spelling I am away from my regular machine, and being unwell is not helping.

I found the dvcodec document really easy, thanks to an open source Quasar DV libdv codec project that gave the link being at the top of Google.

http://www.sony.ca/dvcam/pdfs/dvcam%...20overview.pdf
http://libdv.sourceforge.net/

Reading things and finding out how they work is amazing. When you find out sometimes you think why on earth did they design it thta way. This is such an occasion. DCT in the codec might be suitable to lossy comrpession, but not to lossless comrpession, which we aim for in the projects, putting on massive overheads, There are variouse akternative methods. But enough of that.

Back on subject now. I'll adress this in a way for everybody to read. It uses DCT, but variable length encoding for the compression. DCT is used to isolate and eliminate data, and alot more space to compress more busy areas. Note, they think that low frequemcy data is visually more important than high frequency data, which we want to enhance the footage. So clearly there is preference to low freequency data, that probably explains the blandness in a lot of consumer video. Busy areas are given more distortion, because it is less visible, but in film things are so large it would eb easier to pick up. Also more preference is given to the centre of the picture, where as in film all the picture and the sides, for opposing subjects, is important.

A diiferent quantisastion methord is used for great movement than for stationary objects. You should not note that the break down under complex scenes and movement, should not be as obviouse as HDV when you have more data rate to compensate for lower comrpession. So the panasonic hd camera offers significant advantages. But there still would be reduction ("distortion") in complexity and excessive movement. But to examine it, I suggest you use a big screen upclose, as the "distortion" should be more obviouse at cinema screen sizes. To look for, is a flattening of the image (ie less high res details) to the piont of block and large areas appearing, and reduced res om the sides. Compared to low comrpession this is where a lot fo differences should show. It should even be possible to change the opesource version of the DV codec to favour cinematic concerns, and lower the compression ratio (4:1 or 3:1). I wonder if any of the commercial cameras have special cinema, or action, freindly modes in their codecs that users can switch between. Wouldn't it be funny if those picture modes (like sports mode) do this without us realising it, but I have no idea if they do or not.

I am too sick to finish reading the document, If anybody would like to finish it and confirm all this, that would be great, I got to the vatiable length encoding section.

So the answer is that, yes, if the version of the codec allows it, the letter boxing will help, particularly if the codec frees up space to higher frequency details. You will however, loss the detail gained from the full frames lines being sqeezed down to 16:9, but they were not really seeable on a normal TV anway, and on the large screen they are actually oblong pixels. The question is that can film transfer picture processing and upscaling effectively use this?

Thanks

Wayne.

Keith Wakeham · August 5th, 2005, 05:15 AM

I did a quick skim through and it looks like your right about the dct and vlc. So theoritically it should be possible to reduce artifacts and gain some greater detail.

I don't know about film transfer but It would definetly need some testing to see if the gains are actually useful or if the use of something like a wide screen lens adapetar, (the ones that strect the picture horizontally) would be a more effective way of gaining that superwide format we would want.

Juan M. M. Fiebelkorn · August 5th, 2005, 07:58 PM

if you have a native 16:9 camera, just adding an optical 16:9 adapter to it will give you a final result of around 2.35:1....
Anyway best results would be obtained using something that shoots progressive.

Wayne Morellini · August 5th, 2005, 10:14 PM

Absolutely right, I forgot to mention that. There is a difference in the mechanism to handle interlace movement, so I imagine the normal mechanism is more efficient and used as standard in progressive. I haven't re-examined the document, but the VLC appears to work on areas of two dimensions, if that is so then the progressive should work even better.

Something I noticed, is that it estimates the DCT detail elimination to fit in a 25Mb/s data stream, but if it is wrong and the next frames data rate is under 25Mb/s (I don't know if it averages it down to make room) it tries to store the excess data on unused portions of the tape for the next frame. This means very little to us, basically either the complex movement detail will deteriorate a few frames more slowly due to spread, or the first and last frames will have more detail, with the occasional peaks of detail in between, as the picture allows it. But this gets me to thinking about the 24fps DVX100, what does it do with the extra 6 frames on tape. As I understand it, it records 24 frames per second onto a 30fps tape format. Can the DVX spread the data from each 24fps frame to the other frames, or every 4th 24fps frame, or are they left blank?

Re-edit:
This is ridiculous, I feel better today and I'm on my normal computer, and the system won't let me go back and re-edit those spelling mistakes from yesterday. I think it locks your post once you make another in the same thread. I often fix a post up if I think it was written a bit badly.

Steve Brady · August 6th, 2005, 12:40 AM

Quote:

Originally Posted by Wayne Morellini

But this gets me to thinking about the 24fps DVX100, what does it do with the extra 6 frames on tape. As I understand it, it records 24 frames per second onto a 30fps tape format. Can the DVX spread the data from each 24fps frame to the other frames, or every 4th 24fps frame, or are they left blank?

The middle frame in each sequence of five video frame contains one field from the second 24p frame and one field from the third 24p frame of the corresponding sequence of four 24p frames.

That's possibly a slightly confusing way of putting it; it's easier if you think of fields: You get two fields of the first 24p frame (=one video frame), 3 fields of the second 24p frame (=one and a half video frames), 3 fields of the third 24p frame (=one and a half video frames), then two fields of the fourth 24p frame (=one video frame), then the cycle (1+1½+1½+1=5 frames) repeats. Yes, you're effectively wasting 20% of the bandwidth, but there isn't really an alternative method that'll maintain compatibility with existing DV codecs.

Leo Mandy · August 6th, 2005, 01:49 PM

So does anyone have a optical 16:9 adapter to try this?

Wayne Morellini · August 7th, 2005, 01:44 AM

Actually, an optical 16:9 adaptor should just squeezed more into the 16:9 letter boxing, what would have to be done is something like a 35mm SLR lens adaptor with a 1:2.35 mask put on the projection surface to test it.

How do condenser lens adaptors work, do they require the image to be projected on one end of condenser lens, and then narrow it down, or do they take non-divergent light and narrow it down more? If it does work on the projection plane, that would be as good as an normal lens adaptor and more light. There must be some sort of solution in the film camera adaptors, worth looking at. Otherwise the only thing you could do is stick a cut mask out in front of the camera, and pout up with the unfocused edge of the mask.

Wayne Morellini · August 24th, 2005, 02:05 AM

I have just submitted this thread in the HC1 forum, on letter boxing 720p in a 25Mb/s 1080i HDV stream:

http://www.dvinfo.net/conf/showthrea...164#post350164

Thanks

Wayne.

July 29th, 2005, 02:14 AM	#1
Wayne Morellini Inner Circle Join Date: May 2003 Location: Australia Posts: 2,762	Technique to decrease compression ratio in DV footage You wouldn't believe this, I have figured out a way that should lower the compression ratio of Mini DV footage, but I don't know if the camera codecs will allow it. When you use the cropped wide screen feature, it records black bars top and bottom. But black bars should take up very little compression space in any descent compression routine, leaving more for the rest. I did the calculations weeks ago, but for PAL progressive (that should also require less compression space than interlace) you should get significant reduction in the ratio of 25%. At these ratio's a little reduction should matter a lot. Now if we look at Wide screen cinema formats 2.35:1, is around 43% reduction, that puts you into Pro DV ratio territory. You should be able to make up an lens hood to crop it. The compromise is that for cinema projection you loss the extra resolution of a true 16:9 progressive mode, but retain a square pixel. But one of the big problems with DV compression is rapid complex movement (though nothing compared to HDV 1080i) so it might be possible to use cropped 16:9 for special effects (explosions etc) and true 16:9 for everything else. So cropped wide screen is not a complete loss after all. So what do you think? Follow this link on pixel shifting single chip: http://www.dvinfo.net/conf/showthrea...561#post339561

August 5th, 2005, 10:14 PM	#11
Wayne Morellini Inner Circle Join Date: May 2003 Location: Australia Posts: 2,762	Absolutely right, I forgot to mention that. There is a difference in the mechanism to handle interlace movement, so I imagine the normal mechanism is more efficient and used as standard in progressive. I haven't re-examined the document, but the VLC appears to work on areas of two dimensions, if that is so then the progressive should work even better. Something I noticed, is that it estimates the DCT detail elimination to fit in a 25Mb/s data stream, but if it is wrong and the next frames data rate is under 25Mb/s (I don't know if it averages it down to make room) it tries to store the excess data on unused portions of the tape for the next frame. This means very little to us, basically either the complex movement detail will deteriorate a few frames more slowly due to spread, or the first and last frames will have more detail, with the occasional peaks of detail in between, as the picture allows it. But this gets me to thinking about the 24fps DVX100, what does it do with the extra 6 frames on tape. As I understand it, it records 24 frames per second onto a 30fps tape format. Can the DVX spread the data from each 24fps frame to the other frames, or every 4th 24fps frame, or are they left blank? Re-edit: This is ridiculous, I feel better today and I'm on my normal computer, and the system won't let me go back and re-edit those spelling mistakes from yesterday. I think it locks your post once you make another in the same thread. I often fix a post up if I think it was written a bit badly. Last edited by Wayne Morellini; August 5th, 2005 at 10:21 PM. Reason: Editing spelling problem

July 29th, 2005, 04:12 AM	#2
Keith Wakeham Major Player Join Date: Jan 2005 Location: St. John's, NL, Canada Posts: 416	I was of the understanding that mini dv was a constant bitrate dct compression and that under constant bitrate with dct each block has so many values like a pixel would but because of dct the black area would be given the same x of y values so it wouldn't spare up any extra data. The interframe compression like mpeg 4 saves space because it detects no changes between keyframes and doesn't write much data then, while intraframe still needs to write all the black data. I'm no compression expert so I could be way off but wouldn't having black bars for such a wide format result in sevre loss of resolution. You'd be tooping out at like 720 x 200 then with 4:1:1. That is just an estimated pixels number, to lazy to calc the real one for 2.35:1

July 29th, 2005, 08:21 AM	#3
Wayne Morellini Inner Circle Join Date: May 2003 Location: Australia Posts: 2,762	I think it was closer to 240 pixels. That's the downside, but then again it is the same resolution per inch as the horizontal measurement. The way they get the horizontal up is through resolution upscaling, so it is a compromise. We're not talking about HD here, we are talking about basic results and improving detail (unless you have uncompressed component instead) . About the compression data rate being divided up into blocks, that is the sort of thing I was worried about, but if it is really good it should be intelligent enough to identify that all the surrounding blocks are the same, and compress it as an area, and re-distribute the bandwidth to areas with details. If anybody could give links to positively confirm that it works like you said, and does not have area based compression, that would be good. I can't remember (it has been since the 80's/early 90's since I have read about compression) but I thought they used a couple of different methods in DV like compression.

July 29th, 2005, 08:36 AM	#4
Matthew Wauhkonen Major Player Join Date: Apr 2005 Posts: 285	DV frames are indeed self-contained. DV has no problem with explosions, fast movement, etc. beyond motion blur, which would be a problem with uncompressed footage as well. Turn your shutter speed up to 1/1000th of a second and shoot progressive and you'll notice that everything is super sharp in each frame even during fast motion but your footage looks choppy. So that's just completely not an issue with DV. Never has been, never will be. HDV, however, does have a GOP-based codec (MPEG2) so it does struggle with fast motion (and, yes, I've noticed this in action and it can get very bad, although I still think the FX1 is a really cool product). As for the "black bars" there has been a lot of argument over how smart the DV codec is. Since the hardware DV codec is different in different cameras, some have even claimed that some "dumb" cameras devote as much space to the black bars as to the rest of the frame, while some "smarter" cameras devote more to the rest of the frame. I don't buy it. Although there's no proof either way, I think that the entire frame is compressed so that each pixel is given the same amount of attention. This is a real time codec invented like 15-20 years ago. High quality GOP codecs (unlike HDV) take a long time to compress footage on very fast computers. DV compresses on real time in 200 dollar camcorders. Sorry, it's just not that smart. Furthermore, due to the 4:1:1 colorspace, there will always be moer "color" compression artifacts in a letterboxed image than an anamorphic one. I have a dvx100a and always shoot squeeze mode for 16:9 and can say, although only anecdotally, that it looks much better than letterboxing.

July 29th, 2005, 09:04 AM	#5
Wayne Morellini Inner Circle Join Date: May 2003 Location: Australia Posts: 2,762	Comparing HDV codec compression to DV codec compression on a fast computer would give better results. Computer computational performance is a bit misleading, they are very inefficient compared to custom silicon used in cameras like the HC1. If it devotes the same amount to each pixel, that is horrible and the picture detail should suffer badly. What I am saying actually should be relatively little overhead (compared to Mpeg2). So I think it is possible, but only specs on the inner workings of DV would tell. I have read positive reports of problems with DV in complex motion/detail. On a limited set bandwidth codec, it should only be able to keep up with limited amounts of detail, especially if space is alloted tied to basic blocks. So the detail would be wrong, or worse.

July 29th, 2005, 09:38 AM	#6
Keith Wakeham Major Player Join Date: Jan 2005 Location: St. John's, NL, Canada Posts: 416	It would be nice if we could find a way to test this theory. I don't know how though. would be interesting to see anyway. But for hdv it might work because of the GOP interframe structure rather then the intraframe. Problem with that is how do you get a 16:9 camera to shoor 2.35:1.

July 30th, 2005, 07:13 AM	#7
Wayne Morellini Inner Circle Join Date: May 2003 Location: Australia Posts: 2,762	I think, unless somebody comes by and provides the evidence I will have to look it up on the web and find an accurate description. Unfortunately I don't want to put everything else aside for four hours to do it. Maybe tomorrow, or Monday. Only a usefull tip for others.

August 5th, 2005, 05:15 AM	#9
Keith Wakeham Major Player Join Date: Jan 2005 Location: St. John's, NL, Canada Posts: 416	I did a quick skim through and it looks like your right about the dct and vlc. So theoritically it should be possible to reduce artifacts and gain some greater detail. I don't know about film transfer but It would definetly need some testing to see if the gains are actually useful or if the use of something like a wide screen lens adapetar, (the ones that strect the picture horizontally) would be a more effective way of gaining that superwide format we would want.

August 5th, 2005, 07:58 PM	#10
Juan M. M. Fiebelkorn Major Player Join Date: Jun 2004 Location: Buenos Aires , Argentina Posts: 444	if you have a native 16:9 camera, just adding an optical 16:9 adapter to it will give you a final result of around 2.35:1.... Anyway best results would be obtained using something that shoots progressive.

August 6th, 2005, 01:49 PM	#13
Leo Mandy Major Player Join Date: Dec 2004 Posts: 749	So does anyone have a optical 16:9 adapter to try this?

August 7th, 2005, 01:44 AM	#14
Wayne Morellini Inner Circle Join Date: May 2003 Location: Australia Posts: 2,762	Actually, an optical 16:9 adaptor should just squeezed more into the 16:9 letter boxing, what would have to be done is something like a 35mm SLR lens adaptor with a 1:2.35 mask put on the projection surface to test it. How do condenser lens adaptors work, do they require the image to be projected on one end of condenser lens, and then narrow it down, or do they take non-divergent light and narrow it down more? If it does work on the projection plane, that would be as good as an normal lens adaptor and more light. There must be some sort of solution in the film camera adaptors, worth looking at. Otherwise the only thing you could do is stick a cut mask out in front of the camera, and pout up with the unfocused edge of the mask.

August 24th, 2005, 02:05 AM	#15
Wayne Morellini Inner Circle Join Date: May 2003 Location: Australia Posts: 2,762	I have just submitted this thread in the HC1 forum, on letter boxing 720p in a 25Mb/s 1080i HDV stream: http://www.dvinfo.net/conf/showthrea...164#post350164 Thanks Wayne.