Sumix 2/3" 1920x1080 CMOS

[Noah Yuan-Vogel]

looks like sumix has a new video that is 1080p60. thats pretty impressive, for some reason it is mono, but since there is no extra overhead for color capture, i imagine raw capture at the same rate should be possible. they must have their compression scheme working to be able to push 60fps at 1920x1080 over gigE... would be great to get some real updates from them. all of the lapacian color videos still have some odd patterns on them, the pattern is not visible at half res, but at full res, the pattern sure is weird, it's somewhat maze-like. im not sure how else to describe it, those sure are some sharp pictures though.

[Daniel Lipats]

Thanks for the heads up on the videos.

They are indeed sharp, although the patterns are distracting. I found them to be most prominent in the yellow colors. I hope its something they address soon.

[Farhad Towfiq]

Tuning the sensor in this camera has been not so straight forward to get optimal images for various frequencies and illuminations. We are now ready to post more sample videos with 1080p60. We have some difficulties with FPN which is higher comparing to our other cameras with micron sensors. However random noise is not even detectable at unity gain. With extra 15 dB gain we see some noise comparable to that of our lowest noise camera SMX-M72C without gain. This is very encouraging as proves not only the sensor is not noisy but our electronics/PCB design has been satisfactory and changes will not be needed for the time being.
We are considering to move to newer sensor P4562-3T as quickly as possible and limit this version of the camera with P3562 to production of 20-100 cameras. Therefore, this version will be more like a collector item. By the way, the case is made of military corrosion resistant aluminum alloy in one block with only the faces screwed. The inside is fixed so solid that it can take several hundred G shocks with out losing the sensor alignment.

[John Papadopoulos]

The sharpness in images that have not been debayered by a good algorithm is natural. The difference in the levels of neighbouring pixels is always there because each area has different levels in each of the three colors and each pixel can only have one of the three colors. So, the material will have high frequencies that are not present in the real optical image. If you don't debayer at all, the image is even sharper because the bayer grid stands out.

[Noah Yuan-Vogel]

good to hear from you farhad and john. that is one of the most exciting things about that range of altasens sensors, the latitude/low noise.

regarding the bayer artifacts, that is a bit above my head, but the lapacian bayer on the M73 camera seemed quite useable to me, certainly fewer artifacts than what we are seeing with this new camera. anyway im sure sumix knows we are counting on them to make the debayer as clean as possible, it certainly will make a big difference regarding whether the camera appears attractive as a digital cinema camera.

and in terms of build quality, sounds sturdy, my M73 sure is sturdy, but the paintjob seemed pretty iffy and the black coating would rub off on my hand, i hope this has been improved with this camera. the new camera does look better balanced and a bit more substantial, which is good. the other sumix cameras are nice in their simple design, but its not that conducive to attaching to a tripod or rig since there is little surface contact with a tripod plate or rod system plate. this new one looks like it could be better in that respect.

[Seth Kersey]

How much easier is it to debayer a single image, as opposed to a video sequence?

The rose image debayer seemed much cleaner to me, but I am new to digital video and still do not fully understand it all. I do understand that other factors come into play when dealing with motion, but I am just wondering if a different algorithm was used or if it is just that video is harder to work with.

Farhad, Is the difficulty "tuning" you mentioned responsible for some of the horizontal and vertical lines that seem apparent in both the videos and the rose image?

Will the new sensor help with these difficulties at all?

[Farhad Towfiq]

Noah, For this camera we are using special "hard anod" coating that can not easily be scratched. This coating is usually used for gears and moving parts and quite sturdy. We are considering to use this coating for all other cameras also. However for industrial uses the cameras are not touched regularly.

Seth, correct, tunning changes the FPN and not easy to figure out how. New sensor must have reduced FPN. Altasens claims that in Altachrom family sensors the FPN is completely removed. But the sensor with 2/3" format will be only available end of 2008.
All debayering in the sample videos are done real-time. And our engineers insist on optimizing the performance/speed in real time. I think only with heavy statistical processing and only off line we can bring the best RGB images out of the raw data. Algorithm must look at statistical data, image parameters, many reference images for correcting for FPN, etc. In my graduate work (many years ago) I developed some non parametric statistical methods for estimating a desired signal from multidimensional data. For this you need example training data. For example samples of non-moving images at low frequency can be acquired and used as training data for algorithm to match the high frequency images to them.

[John Papadopoulos]

There is little difference in debayering video and still images. There are some issues that have to do with how debayer stands heavy processing (which debayer algorithms are not designed for) but still image quality should look good even in a realtime debayer solution anyway. It all depends on the debayer technology, its efficiency and quality.

This algorithm does more than 400 720p frames per second on an average pc and it doesn't look that bad, does it? It can run multitrack 1080p editing easily.

http://img112.imageshack.us/img112/1435/dbframevb0.jpg

This 2/3" 1080p cmos camera is similarly priced to an 1/2" 720p CCD solution, so if anyone does not need CCD, it could probably be added to our camera sensor options. Does it output uncompressed? The image appears to be very unsaturated in the flower and phone samples though and it does appear to not have very clean shadows. Similar to an old Altasens head I tested.

[Farhad Towfiq]

Reconstructing RGB from Bayer images indeed is a fascinating subject. The more information is preserved in the Bayer images the better the result using an "optimum" algorithm. I would prefer a sensor that even has more FPN but preserves the information rather than a sensor that does not have FPN but also smooths out visual clues that are used by eyes. Early vision processes in eyes before even the transmission to the brain detect such features as edges, motion, 3D from motion, 3D from shading, enhances color and contrast, etc. Therefore a good deBayering algorithm must preserve as much as possible those features used by early vision. Removing of FPN is possible with intensive use of conditional histograms. Any spatial filtering must preserve or even enhance those visual clues used by eyes. There is no easy way but to experiment. First finding out what sensors are more information rich, it helps make our experiments more productive.

[John Papadopoulos]

Even minimal artifacting is distracting though, while a small reduction in test chart performance will only be visible in test charts and only when comparing side by side. A natural image is a lot more important than test charts. I remember the test charts after the introduction of a low resolution HD camera, shapening setting was probably at maximum, there was aliasing all over, the contrast transition profiles were unacceptable due to tons of overshoot, there were 5 pixel halos, and everytbody was arguing about the resolution shown in the charts! A properly implemented imaging system has smooth transition profiles with zero overshoot and is free from artifacts. Film is like that, good scanners and cameras are like that when operating with good settings, and every single scientific imaging system has these qualities.

I believe there is a bug in the debayer implementation in the sample with the mobile phone. It shouldn't look like that. This is not sharpness, this is 100% artifacts. Try creating a bayer image out of a high quality test image by keeping only the color info a bayer sensor would get, debayer with that algorithm and compare with the original image. It's not detail, it's artifacts that are not present in the original optical image. I have seen implementation errors before, the Fire-I app I tried when we were starting this project has a bug in the blue color debayer compared to any other bilinear implementation and still nobody noticed:)

[John Papadopoulos]

There is also a bug in the debayer on the color 1280x720 video sample smx-11m5c(8)-milk-color.zip. The green resolution is half of what is should be horizontally. A bilinear debayer on the raw video looks fine (on the left).

http://img253.imageshack.us/img253/5...kbayer2wm2.jpg

[Farhad Towfiq]

John, Thank you for pointing out the problem. This is a serious bug that passed by our testers.
About Fixed pattern noise and other artifacts like aliasing, I agree that they are not acceptable specially in moving images as artifacts may move with different speed or even in the opposite direction as it is possible with aliasing. Artifact due to lossy compression and decompression are just awful. The fixed pattern noise as they can be predicted can be removed given we can collect enough samples. Correlated noise in principal can be reduced and made random, the same is true for some convolution artifacts like shifted images if there is a sharp main lobe. In any case a clean sensor is the best. But a low noise and high dynamic range sensor with removable artifacts and enhanceable RGB will be more desirable than an artifact clean sensor with more random noise.

[Paul Curtis]

This new GIGe camera has appealed to the innergeek within me. I'd love to find out a bit more about this camera and in general. I hope this is the right thread to ask!

Farhad i assume you work for sumix? Perhaps you can advise me on:

are there heat problems with the casing? (there doesn't seem to be ventilation and i thought sensors get quite hot)

Would a laptop work as a capture station, what are the real minimum specs? (there's a wealth of small form factors PCs and ultra portables out there).

What is the data rate down that GigE?

Why don't more people use cmount lenses? What's the problem with them!? (because there seems to be some great product out there). Are the focus ranges too short for filmmaking (i know the lenses are physically small)? Are the lenses good because they're only just 2/3rds, have a short flange depth and are simple? (i've seen f0.95 primes around). For example the SI 2K camera is all PL mount, is that just to conform to the industry or are their optical issues to think about. (im thinking about the problems leica had with using their lenses on digital because of the microlenses and short depth)

What are the realistic time scales for these being readily available? (especially with the new altasens sensor - the one in the SI 2K)?

Is the RAW data from the head standardised, so if a new head came along the software would be the same?

What are the real issues in pulling something together that works in the field?

cheers
paul

[John Papadopoulos]

While Farhad answers, I will add some info on lenses.

C mount is just a mount. It keeps the lens at a distance from the sensor, just like any other mount. If you use a C-mount to PL of F-mount adapter, the system is identical to using a PL mount directly. There is nothing in the light path, you just add some metal to move the lens further away and a mechanical system to properly attach the new lens. The sensor might be smaller than the format of the lens, but that has nothing to do with the mount.

C-Mount lenses come in many flavours. Cine lenses, tv lenses, security lenses and high quality machine vision lenses. The cine and machine vision type are usually sharp enough with low distortion. The cine lenses are softer than good machine vision primes though, especially at large apertures. The tv lens category is very sd in resolution and usually comes with lots of distortion, I mean LOTS of barrel distortion. The security/cctv lenses are usually junk and this is the usual c-mount lens type.

The cost difference has to do with features and manufacturing requirements. A $300 lens that has a motor version can cost $2000, a zoom has to cost 5x or 20x in order to keep a little of the prime performance. The basic c-mount megapixel prime is quite good. The mechanics do not help for pulling focus or any focus addons because the barrel is quite small in diameter (3 to 5cm) and there is no room to add anything or even position your fingers really. You can use these lenses and their optical performance is excellent because they are modern designs, but the manual handling is problematic. You can expect a good megapixel c-mount prime to outperform a $30,000 zoom lens or any bundled zoom lens. The design is simple and it's a prime so it's easy to get right and the market is quite large and competitive. You will not get 5pixels of chromatic abberations with a machine vision prime. Zoom lenses is not something that is normally required in machine vision, so there are not many of those. There is a large selection of very high quality macro lenses if you find that interesting.

You can always find a normal cheap (non megapixel) c-mount or even c-mount zoom that performs quite well, but chances are small.

Old cheap c-mount cine lenses can look great, but don't expect much sharpness or multicoating or anything like that. You might like the look though and the better manual control.

The reasonable choice is selected c-mount megapixel primes if the shot can take it, or PL mount lenses with an adapter. The professional choice is PL mount lenses. If the pixels are large and you don't need wide angle, you can use f-mount lenses with an adapter and a large crop factor depending on the sensor size. You can also use a DOF adapter.

[Paul Curtis]

John,

Thank you for your information (i've also found your other thread very interesting).

There seems to be some very nice schneider c mount lenses, $1500 range, going as low as f0.95 in some cases (17mm f0.95!). And as you say all the way down to a few hundred bucks. It's difficult to tell sometimes whether manufacturers specify the focal length for a particular sensor size or for 35mm..

Physically they are small, some seem to have little handles on the barrels to help with access. Do you happen to know how much travel the focus rings have? On most modern SLR the travel length is tiny because it's all autofocus where as cine lenses are much longer.

The confusion i have though is that they are so small compared to other lenses. I mean i have a 10-22 canon for the APS-S sized sensor and it strikes me as much larger than the cmount for 2/3 or 1in sensors. I just wonder if there's some optical math here im missing!

Are machine vision primes in c mount an undiscovered(ish) treasure for cinematography?

cheers
paul