[falconeye]

I create this thread to discuss two topics:

- Rumors and news about video specs
- Principle problems to have decent video with a DSLR (my first post below)

(I started discussion about the latter in another thread and moved it here for further discussion).


I will edit this first post to reflect what we know (in majority agree) about the K-7 video mode:

K-7 video mode specs
(I mark with "#*#" if its the common speculation, not a fact yet)

• Video mode
• Selection of 3 video modes 640x416 (3:2), 1280x720 ("HD" 16:9), 1536x1024 (3:2).
• 30fps in all three modes
• Recording limit is 4 GB or 25 min per movie clip. (The 25 min limit may be fantasy for EU officials).
• MJPEG codec in AVI container 74 MBit/s (1024p 3:2) and 58 MBit/s (720p)
  MJPEG codec delivers still image quality (300 kByte per frame) and no motion compression artifacts.
• Contrast AF in video mode before start of capture, not during capture (with firmware v1.00, CAF during capture was only possible up to v0.20).
• Aperture control is possible in movie mode (before start of capture in A aperture, during capture with aperture ring.)
• Exposure is automatic but can be locked with AEL prior to a take.
• SR is active during movie mode.
• Exposure compensation is possible during movie mode.
• LV menu option: "Movie Overlay"
• Mono microphone on body
• Microphone jack
• No recording video out over HDMI
• LV out over HDMI (confirmed!)
• LV out over AV out
• AV out, supporting SDTV
• HDMI out, supporting SDTV and 720p and 1080i
• Video frames are sub-sampled from the full sensor data, where each RGB video pixel is constructed from 3 out of the 18 sensor cells lying within a 6(horizontal) x 3(vertical) array (verified for firmware 0.34, very likely to be so in final firmware).
• So, the first processing stage gives two similiar sets of 768 x 1024 RGB pixels, which is then scaled to 1536x1024 w/o further interpolation.
• The 720p video resolution is supersampled from the 1536x1024 resolution.
• The lack of interpolation leads to aliasing artifacts at the full 1536x1024 resolution, not visible at 720p.
• The choice of the wide subsampling matrix may lead to purple and green fringing of two pixels width (1.5px at 720p) to both sides of a vertical high contrast border. This effect is much less visible at 720p.
• Using every 6th pixel leads to a low light performance between camcorders and the Canon 5DmkII or Red ONE.
• There are no other artifacts than those described above. The video images are very smooth and compression artifact free.

---

Last edited by falconeye; 06-22-2009 at 03:16 AM.

[falconeye]

Is only a (sparse) subset of the sensor surface used in DSLR video mode?

I fear that exactly this may be the case. Read here:

HD data is 40 megabits per second which translates to 5 or 6 megabytes per second of material to be transfered to the SD. However this is the bandwith AFTER processing and compression. The RAW data being output by the CMOS is significantly bigger (I don't know by what factor).

Interesting math here.
BTW, HD camcorders record at much less than 40 MBit/s (less than half of this, actually). Only 5dmkii comes close. 40 MBit/s is the (max.) Bluray spec for the HD video part of a data stream.

First, and most importantly, we will have to know, if:

1920x1080@24fps is achieved by:
- subsampling (only 1/7 of pixels are read out): 142 MByte/s input rate (8 Bit RGB)
- supersampling (all pixels are read out, demosaiced in the trivial manner and supersampled to half size): 564 MByte/s input rate (16 Bit Raw) and (with ~4 instructions to supersample a pixel) about 0.3 GOps/s processing power to condense the input into a 211 MByte/s input rate (8 Bit RGB). To further reduce size and compress into H.264 is downhill from there (where realtime H.264 compression is requiring hardware acceleration, of course).

BTW, 3.5 fps shooting requires 97 MByte/s input rate (16 Bit Raw) and (with ~20 instructions to demosaice a pixel) 1 GOps/s processing power. The latter is not available which means that the buffer fills up faster than it can be written out in a burst. Maintainable speed suggests that about 0.5 GOps/s are available.

So, we see that supersampling is possible but requires ~5 times the speed of internal data busses compared to subsampling. Or it requires a so-called hardware-binning feature (or pixel binning, which can be done on the CMOS chip). Then, an increase by ~2 times the speed would suffice. Samsung did announce a CMOS chip with pixel binning in December. But only 720p for web cams


WHAT DOES IT ALL MEAN?

Well, most likely, the K-7 features 1080p@24fps HD video but NOT with the enhanced noise/DR performance as known from still photography. It features the DoF control but dismisses the superior IQ which would be possible as well.

There is a slight chance that Pentax increased the internal busses by 5x (or included hardware binning).

If they did, they may smash the entire competition for high end video. But I doubt so

[falconeye]

Philippos answered:

Supersampling a 15.1 MP sensor to 1080p would probably create much heat on the sensor and create too much unnecessarynoise, as the sensor would have to output huge amounts of data for a long time.
I do not know how the current burst mode in the K20D works, but I believe not all photodiodes on the chip function when recording. It's the most efficient way to go, both in terms of computing power needed but most important, it requires far less power.
[source: https://www.pentaxforums.com/forums/577656-post38.html]

[falconeye]

Supersampling a 15.1 MP sensor to 1080p would probably create much heat [...]

So, you argue in favour of my option #1 (subsampling).

Note that pixel binning on the CMOS sensor chip would solve the problems you mentioned. But anyway.

If subsampling is used, a DSLR WITH AS FEW AS POSSIBLE pixels would render the best possible HD video!

With only 1/7th of pixels read out (14.6MP), an APS-C sensor effectively becomes

a 1/2.3" sensor

in video mode (not for DoF -- but for ISO performance)! This is small P&S size.


Sincerely, I must say, that a DSLR video mode capturing only as much light as a camcorder, ie., not exploiting its expensive lenses, would disappoint me.

So, probably, all available DSLR video implementations, K-7 included, would disappoint me.

Anybody knowing anything more specific about this issue? Never heard about it before. This is born out of my own reasoning.


P.S.
Where is your 15.1 MP spec coming from?

[two68s]

I personally don't think HD video belongs on a DSLR. To me, it's a sales ploy for NIKANON. If I want to shoot HD video I'll get an HD video recorder.

[falconeye]

I personally don't think HD video belongs on a DSLR. To me, it's a sales ploy for NIKANON. If I want to shoot HD video I'll get an HD video recorder.

I see your point and partly agree.

One of the reasons why I never was interested in video was the bad image quality, even with current HD camcorders. Many still image photographers may actually feel the same.

BUT

Video from a DSLR could change this. IF all pixels are used. Why I started this discussion...

[Urkeldaedalus]

After using the Canon G9 and G10 for some online video journalism projects, I really think that the video quality standards should be broken down into two categories -- those that can be viewed well on a large screen, and those that are good enough online.

Personally, since I will likely use the K-7s videos for similar (web-based) applications, my concerns are more with the sound and smoothness for the video than necessarily with the resolution or picture quality.

I thought this was something to bear in mind, because what might be disappointing for Indie Filmmakers (anything less than full HD) might be fine for the vast majority of consumers.

[rparmar]

Video is very useful for journalists in the field who can grab some motion stuff to go with the images at no additional cost to their employers. Of course the one factor seriously lacking in this is the audio, so it's best if it's used with v/o or as silent stock.

But of course that is not a real market for Pentax. Instead, video is just a shiny new gimmick toy, and somewhat out of line for what I expected from them.

That said, if it adds features without adding cost or breaking ergonomics, why should I care?

[falconeye]

-- those that can be viewed well on a large screen, and those that are good enough online.
[...]
my concerns are more with the sound and smoothness for the video
[...]
what might be disappointing for Indie Filmmakers (anything less than full HD) might be fine for the vast majority of consumers.

Thank you for enlarging my perspective. You are right.

[kenyee]

Is only a (sparse) subset of the sensor surface used in DSLR video mode?

AFAIK, none of the current DSLR video solutions use the entire sensor surface. They'd need a lot of data xfer bandwidth than they have now..

[konraDarnok]

dSLR video is just starting to change the world of consumer HD video. With their larger sensors and shallow depth of field, dSLR's can produce cinematic video unrivaled by any handheld camcorder. That said, dSLR's have some serious limitations with regard to video capture

[konraDarnok]

AFAIK, none of the current DSLR video solutions use the entire sensor surface. They'd need a lot of data xfer bandwidth than they have now..

And why would it need to? A HD display is only 2.1 megapixel. And frankly, most of my video needs are good at 640x480.

It would be really cool if it had a high speed frame rate tho, even at 320x240.

[farfisa]

From what I understand (ok, not too much at this point), but the first implementation of this (on the D90) was just taking the live view "feed" and recording it.

Thus it would use the entire surface and down-sample.

I'm not sure how HD video cameras do it either--they all seem to have sensors with more pixels than they need. The latest Canon HF S10, for example features a 1/2.6-inch CMOS sensor with 8.59 Megapixels. It uses 6 MP for video (according to camcorderinfo.com), but 1920 x 1080 = 2 MP.

I can't see them having a mode that can project the image on to a smaller part of the sensor surface--it's easier (because it's already being done for live view) just to take the full surface and downsample.

Now how this is done and how fast it can refresh--that'll determine how useable the video is (wrt motion artifacts).

My guess--it'll be 720p, just based on the 21 fps burst mode of the K20D--they could crop what they already have (1536x1024) to 1280 x 720 and hopefully get that frame rate up to 24.

Again, I'm not sure about any of this, but if you think of the logistics of isolating the center 1920x1080 pixels of the sensor through cropping, or some other projection vs. downsampling, I think it makes more sense to downsample.

[Urkeldaedalus]

Thank you for enlarging my perspective. You are right.

Thanks to you for bringing up the technical aspect of the video, which I am admittedly not very versed on at all. (I'm more qualitative than quantitative when it comes to video at this point).

From my perspective I am excited about video in the Pentax because when I was asked to do video on one of my assignments, I was also often asked to do photos and reporting as well -- an increasing trend in journalism. And as light as the G9 and K20D combo were, the less things I have on me to carry (or potentially lose) when I am out reporting on something the better.

As for whether this will open a new market for Pentax with Photojournalists, I kind of doubt it, as most (American at least) newspaper photogs are firmly Canon loyalists from what I've noticed. Not that I blame them, it's just a tough market for Pentax to break into.

[yeedub]

I personally don't think HD video belongs on a DSLR. To me, it's a sales ploy for NIKANON. If I want to shoot HD video I'll get an HD video recorder.

I like the video feature; you cannot get the same depth of field with a consumer HD video camcorder that you can with an SLR and lens.