4:4:4 10bit single CMOS HD project

Rob Scott · July 27th, 2004, 12:33 PM

Quote:

Ben Syverson wrote:
24p is one of the last remaining things they won't put in inexpensive cameras. I think the fear is that people will flock to the cheap stuff rather than the high-margin expensive stuff.

Yup, I know it. It's like IBM in the early days of the PC -- they didn't want to add advanced features because they didn't want to cannibalize their higher-end market. But then other companies cannibalized it anyway.

Aaron Shaw · July 27th, 2004, 12:36 PM

What sort of data transfer rate would be required to transmit 4:2:2 footage to harddisk?

Juan M. M. Fiebelkorn · July 27th, 2004, 12:44 PM

10 bit data, 4:2:2 for 1920x1080 takes @24 fps takes 119 MB/s.

Two WD Raptor disks in RAID 0 are enough to sustain that transfer rate.Although I think you would need a motherboard with a data bus higher than 133 MB.

Does anybody know about someplace to look at image test samples from the Altasens chip??

My main goal is to make a 2K or 1920x1080 camera, and up to now I'm heading for the SI-3170 which uses a chip called YM-3170 from a company called Y-Media situated in California which doesn't have a Web page.

Anyway it seems to me that this is a rather old chip.It was presented September 2000, so I don't know.
More data welcome.....

Aaron Shaw · July 27th, 2004, 12:49 PM

What about 12bit 720p 60fps (ie the JVC camera)? How does one calculate this? I've never needed to do it before.

Ben Syverson · July 27th, 2004, 01:27 PM

@Juan: 10 bit data, 4:2:2 for 1920x1080 takes @24 fps takes 119 MB/s.

True, but remember that we're transferring and storing 10 bit monochrome data from the sensor (ie, before it's de-Bayered into RGB data).

So to calculate 4:4:4 for 10 bit at 1920x1080 at 24fps, we do this (this works for any Bayer calculation):

1920 * 1080 * 10bit / 8bits-per-byte == 2592000 bytes per frame
2592000 * 24 = 62208000 bytes per second.
62208000 / 1024 = 60750 kilobytes per second.
60750 / 1024 = 59.326 megabytes per second

So under 60 MB/sec for uncompressed 4:4:4 10-bit 1080p @ 24fps. If we add lossless compression, we can probably cut that almost in half.

4:2:2 is not a good idea for transfer and storage of Bayer-filter data. It's a false economy, because 4:2:2 refers to luma and chroma. Those terms only work if you're dealing with YUV/YCbCr or RGB. So you wind up with twice as much data for 4:2:2 than 4:4:4 uncompressed. Here's the calculation for 4:2:2 10bit 1080p @ 24fps:

1920 * 1080 * 10 / 8bits == 2592000 bytes for the luma channel
1920 * 1080 * 10 / 8bits == 2592000 bytes for the two chroma channels
(each chroma channel is half-size, so together, they equal one full-size frame)
5184000 bytes per frame, 124416000 bytes per second
124416000 / 1024 / 1024 = 118.65 MB/sec.

4:2:2 takes twice the data to store and transfer, with no benefits. 4:4:4 is the way to go.

By the way, I talked to Sumix about a FPS lock, to which Farhad responded "We are implementing exactly this."

And Sumix seems to be very open to implementing a Cineon exporter in their 10-bit software update. That would rock.

- ben

Juan M. M. Fiebelkorn · July 27th, 2004, 01:57 PM

Does Sumix have support for RAW bayer data, or it only outputs RGB???

Will they give us some kind of open source applet or an SDK???

Have you thought about using a 16x9 adapter to use all of the sensor´s area to get a better image???

Does anybody know about someplace to look at image test samples from the Altasens chip??

My main goal is to make a 2K or 1920x1080 camera, and up to now I'm heading for the SI-3170 which uses a chip called YM-3170 from a company called Y-Media situated in California which doesn't have a Web page.

Anyway it seems to me that this is a rather old chip.It was presented September 2000, so I don't know.
More info welcome.....

Ben Syverson · July 27th, 2004, 02:02 PM

@Juan: Does Sumix have support for RAW bayer data, or it only outputs RGB???

You can output RAW B&W bayer data or de-Bayered RGB data via nearest-neighbor, bilinear or laplacian de-mosaicing algorithms. I only output RAW.

@Juan: Have you thought about using a 16x9 adapter to use all of the sensor´s area to get a better image???

Absolutely. I have a 16:9 adapter and a 2.33:1 adapter I'll be testing with in the coming weeks and months.

- ben

Juan M. M. Fiebelkorn · July 27th, 2004, 02:05 PM

GREAT!!! thank you :).

I believe that getting the image thru anamorphic, be-bayer it, and resize to 1280x720 would give amazing quality!! (in fact almost 4:4:4!! You will be getting 512 instead of 360 lines for Red and Blue)

Will they give us some kind of open source applet or an SDK???

Ben Syverson · July 27th, 2004, 02:07 PM

Definitely! My thinking is that if you do anamorphic 1280x1024, you should be able to expand it to 1920x1080 and have it look okay too!

Juan M. M. Fiebelkorn · July 27th, 2004, 02:11 PM

Sure!!!

Anyway in that situation I would prefer using a 1920 bayer.

My idea is using that 2048x1536 chip with anamorphic to get a final 2048x1107 image that will rock HDCAM!!!!

Or going to 1920x1080 for TV and get better results!!
(Remember that in fact HDCAM is 1440x1080)

@OBIN

Could you give us an exposure test of your camera. I´d prefer an outdoors image on a sunny day with some deep shadows somewhere.
That way we would be able to determine the camera characteristics...
Also it would be nice if you could tell us at what aperture your problems with streaks start...and if they disappear at lower apertures or remain the same...

Ben Syverson · July 27th, 2004, 02:25 PM

@Juan: Remember that in fact HDCAM is 1440x1080

And that's just luminance -- the chroma is 480x1080! That's essentially 3:1:1. And it's highly, highly compressed -- MiniDV has 5:1 compression, but at 135megabits/sec (only 16.8Megabytes/sec), HDCAM is 8.4:1 compression.

Less color resolution than DV, and more compression? I think I'll stick with our 4:4:4 10bit uncompressed, at 1/100th the cost! :)

- ben

Obin Olson · July 27th, 2004, 02:26 PM

I am not at work today but I am going down for a few min to pickup the camera and bring it home for a few TESTS

I may get the SI-3300 camera for testing soon ..less streaking I hope ;)

Juan M. M. Fiebelkorn · July 27th, 2004, 02:30 PM

OBIN,

Is it the 3300 or the 3170???
Is it the 3300 2048x1536?
Does it support 30 fps at full resolution??

Do you have a photometer?
If so, could you measure its sensitivity too?

What is your usual gamma correction??

I guess somebody at SI should make these test for us, but.......

Obin Olson · July 27th, 2004, 04:43 PM

Anyone who wants more test images with HIGHCONTRAST stuff please provide your email...I can't upload ftp from home (don't know login)

I have been fooling around and it turns out that yes they more you GAIN the camera the worse the "smear" is...I now question if I had gain REALLY high on the shoot with the backlights....it was so much pressure I could have overlooked that I think...it's weird ...testing at home here I have to push it HARD to get that amount os smear again...

Steve Nordhauser · July 27th, 2004, 05:38 PM

I'm baaaackkkk.
I've also read this thread and caught up - bouncing up and down in my chair wanting to answer things but I figured that I should read to the end first. So:

Juan: The SI-3170 has been correctly stated as the Y-Media chip - very fast (100Mpix/sec) 3.2 mpix. It uses two 12 bit A/D converters but has a fair amount of FPN and low sensitivity. I put it on par with the IBIS-5. I wouldn't suggest it for studio work with the Altasens coming out. But hey, I wouldn't suggest the IBIS-5 either and we have an external A/D for lower noise and higher data rates. If someone wants to try it on camera link at 12 bit, let me know. The SI-3300 is a Micron camera - 3.2Mpix at 12fps but 1920x1080@24fps.

More on IBIS-5: Yes you can use it in global shutter mode for continuous capture. The integration time is sequential to the readout time so every time you need to change exposure you must recalculate the pixel clock to get the frame rate you want. And, the readout time must be fast enough to handle that. This is probably why Sumix doesn't do it on USB.

Jason: I think Rob explained the 10/12 bits on the SI-1300. It is a 10 bit A/D with the data shifted up by two. This is because most industrial apps only know 8 and 12 bit data.

Obin and all on smearing: Obin, try running at only 24fps instead of 48. Micron suggests lower clock rates for less smearing. This is the rate the IBIS-5 is being tested at anyway. I've tried to be open about this as soon as we heard about the problem. I'm also letting anyone who has an SI-1300 upgrade to either the SI-3300 or SI-1920HD if they need to (the Altasens is more money).

Here is a link for an Altasens monochrome 8 bit image to get everyone thinking:
http://host196.ipowerweb.com/~silico...20in%20HD3.jpg

July 27th, 2004, 05:38 PM	#1140
Steve Nordhauser Silicon Imaging, Inc. Join Date: May 2004 Location: Troy, NY USA Posts: 325	I'm baaaackkkk. I've also read this thread and caught up - bouncing up and down in my chair wanting to answer things but I figured that I should read to the end first. So: Juan: The SI-3170 has been correctly stated as the Y-Media chip - very fast (100Mpix/sec) 3.2 mpix. It uses two 12 bit A/D converters but has a fair amount of FPN and low sensitivity. I put it on par with the IBIS-5. I wouldn't suggest it for studio work with the Altasens coming out. But hey, I wouldn't suggest the IBIS-5 either and we have an external A/D for lower noise and higher data rates. If someone wants to try it on camera link at 12 bit, let me know. The SI-3300 is a Micron camera - 3.2Mpix at 12fps but 1920x1080@24fps. More on IBIS-5: Yes you can use it in global shutter mode for continuous capture. The integration time is sequential to the readout time so every time you need to change exposure you must recalculate the pixel clock to get the frame rate you want. And, the readout time must be fast enough to handle that. This is probably why Sumix doesn't do it on USB. Jason: I think Rob explained the 10/12 bits on the SI-1300. It is a 10 bit A/D with the data shifted up by two. This is because most industrial apps only know 8 and 12 bit data. Obin and all on smearing: Obin, try running at only 24fps instead of 48. Micron suggests lower clock rates for less smearing. This is the rate the IBIS-5 is being tested at anyway. I've tried to be open about this as soon as we heard about the problem. I'm also letting anyone who has an SI-1300 upgrade to either the SI-3300 or SI-1920HD if they need to (the Altasens is more money). Here is a link for an Altasens monochrome 8 bit image to get everyone thinking: http://host196.ipowerweb.com/~silico...20in%20HD3.jpg __________________ Silicon Imaging, Inc. We see the Light! http://www.siliconimaging.com

July 27th, 2004, 12:36 PM	#1127
Aaron Shaw Major Player Join Date: Jul 2004 Posts: 587	What sort of data transfer rate would be required to transmit 4:2:2 footage to harddisk?

July 27th, 2004, 12:44 PM	#1128
Juan M. M. Fiebelkorn Major Player Join Date: Jun 2004 Location: Buenos Aires , Argentina Posts: 444	10 bit data, 4:2:2 for 1920x1080 takes @24 fps takes 119 MB/s. Two WD Raptor disks in RAID 0 are enough to sustain that transfer rate.Although I think you would need a motherboard with a data bus higher than 133 MB. Does anybody know about someplace to look at image test samples from the Altasens chip?? My main goal is to make a 2K or 1920x1080 camera, and up to now I'm heading for the SI-3170 which uses a chip called YM-3170 from a company called Y-Media situated in California which doesn't have a Web page. Anyway it seems to me that this is a rather old chip.It was presented September 2000, so I don't know. More data welcome.....

July 27th, 2004, 12:49 PM	#1129
Aaron Shaw Major Player Join Date: Jul 2004 Posts: 587	What about 12bit 720p 60fps (ie the JVC camera)? How does one calculate this? I've never needed to do it before.

July 27th, 2004, 01:27 PM	#1130
Ben Syverson Major Player Join Date: Mar 2004 Location: chicago Posts: 434	@Juan: 10 bit data, 4:2:2 for 1920x1080 takes @24 fps takes 119 MB/s. True, but remember that we're transferring and storing 10 bit monochrome data from the sensor (ie, before it's de-Bayered into RGB data). So to calculate 4:4:4 for 10 bit at 1920x1080 at 24fps, we do this (this works for any Bayer calculation): 1920 * 1080 * 10bit / 8bits-per-byte == 2592000 bytes per frame 2592000 * 24 = 62208000 bytes per second. 62208000 / 1024 = 60750 kilobytes per second. 60750 / 1024 = 59.326 megabytes per second So under 60 MB/sec for uncompressed 4:4:4 10-bit 1080p @ 24fps. If we add lossless compression, we can probably cut that almost in half. 4:2:2 is not a good idea for transfer and storage of Bayer-filter data. It's a false economy, because 4:2:2 refers to luma and chroma. Those terms only work if you're dealing with YUV/YCbCr or RGB. *So you wind up with twice* as much data for 4:2:2 than 4:4:4 uncompressed.** Here's the calculation for 4:2:2 10bit 1080p @ 24fps: 1920 * 1080 * 10 / 8bits == 2592000 bytes for the luma channel 1920 * 1080 * 10 / 8bits == 2592000 bytes for the two chroma channels (each chroma channel is half-size, so together, they equal one full-size frame) 5184000 bytes per frame, 124416000 bytes per second 124416000 / 1024 / 1024 = 118.65 MB/sec. 4:2:2 takes twice the data to store and transfer, with no benefits. 4:4:4 is the way to go. By the way, I talked to Sumix about a FPS lock, to which Farhad responded "We are implementing exactly this." And Sumix seems to be very open to implementing a Cineon exporter in their 10-bit software update. That would rock. - ben

July 27th, 2004, 01:57 PM	#1131
Juan M. M. Fiebelkorn Major Player Join Date: Jun 2004 Location: Buenos Aires , Argentina Posts: 444	Does Sumix have support for RAW bayer data, or it only outputs RGB??? Will they give us some kind of open source applet or an SDK??? Have you thought about using a 16x9 adapter to use all of the sensor´s area to get a better image??? Does anybody know about someplace to look at image test samples from the Altasens chip?? My main goal is to make a 2K or 1920x1080 camera, and up to now I'm heading for the SI-3170 which uses a chip called YM-3170 from a company called Y-Media situated in California which doesn't have a Web page. Anyway it seems to me that this is a rather old chip.It was presented September 2000, so I don't know. More info welcome.....

July 27th, 2004, 02:02 PM	#1132
Ben Syverson Major Player Join Date: Mar 2004 Location: chicago Posts: 434	@Juan: Does Sumix have support for RAW bayer data, or it only outputs RGB??? You can output RAW B&W bayer data or de-Bayered RGB data via nearest-neighbor, bilinear or laplacian de-mosaicing algorithms. I only output RAW. @Juan: Have you thought about using a 16x9 adapter to use all of the sensor´s area to get a better image??? Absolutely. I have a 16:9 adapter and a 2.33:1 adapter I'll be testing with in the coming weeks and months. - ben

July 27th, 2004, 02:05 PM	#1133
Juan M. M. Fiebelkorn Major Player Join Date: Jun 2004 Location: Buenos Aires , Argentina Posts: 444	GREAT!!! thank you :). I believe that getting the image thru anamorphic, be-bayer it, and resize to 1280x720 would give amazing quality!! (in fact almost 4:4:4!! You will be getting 512 instead of 360 lines for Red and Blue) Will they give us some kind of open source applet or an SDK???

July 27th, 2004, 02:07 PM	#1134
Ben Syverson Major Player Join Date: Mar 2004 Location: chicago Posts: 434	Definitely! My thinking is that if you do anamorphic 1280x1024, you should be able to expand it to 1920x1080 and have it look okay too!

July 27th, 2004, 02:11 PM	#1135
Juan M. M. Fiebelkorn Major Player Join Date: Jun 2004 Location: Buenos Aires , Argentina Posts: 444	Sure!!! Anyway in that situation I would prefer using a 1920 bayer. My idea is using that 2048x1536 chip with anamorphic to get a final 2048x1107 image that will rock HDCAM!!!! Or going to 1920x1080 for TV and get better results!! (Remember that in fact HDCAM is 1440x1080) @OBIN Could you give us an exposure test of your camera. I´d prefer an outdoors image on a sunny day with some deep shadows somewhere. That way we would be able to determine the camera characteristics... Also it would be nice if you could tell us at what aperture your problems with streaks start...and if they disappear at lower apertures or remain the same...

July 27th, 2004, 02:25 PM	#1136
Ben Syverson Major Player Join Date: Mar 2004 Location: chicago Posts: 434	@Juan: Remember that in fact HDCAM is 1440x1080 And that's just luminance -- the chroma is 480x1080! That's essentially 3:1:1. And it's highly, highly compressed -- MiniDV has 5:1 compression, but at 135megabits/sec (only 16.8Megabytes/sec), HDCAM is 8.4:1 compression. Less color resolution than DV, and more compression? I think I'll stick with our 4:4:4 10bit uncompressed, at 1/100th the cost! :) - ben

July 27th, 2004, 02:26 PM	#1137
Obin Olson Trustee Join Date: Jan 2003 Location: Wilmington NC Posts: 1,414	I am not at work today but I am going down for a few min to pickup the camera and bring it home for a few TESTS I may get the SI-3300 camera for testing soon ..less streaking I hope ;)

July 27th, 2004, 02:30 PM	#1138
Juan M. M. Fiebelkorn Major Player Join Date: Jun 2004 Location: Buenos Aires , Argentina Posts: 444	OBIN, Is it the 3300 or the 3170??? Is it the 3300 2048x1536? Does it support 30 fps at full resolution?? Do you have a photometer? If so, could you measure its sensitivity too? What is your usual gamma correction?? I guess somebody at SI should make these test for us, but.......

July 27th, 2004, 04:43 PM	#1139
Obin Olson Trustee Join Date: Jan 2003 Location: Wilmington NC Posts: 1,414	Anyone who wants more test images with HIGHCONTRAST stuff please provide your email...I can't upload ftp from home (don't know login) I have been fooling around and it turns out that yes they more you GAIN the camera the worse the "smear" is...I now question if I had gain REALLY high on the shoot with the backlights....it was so much pressure I could have overlooked that I think...it's weird ...testing at home here I have to push it HARD to get that amount os smear again...