4MP Low Light Weakness Tested 2016

Geutebruck USA • IPVMU Certified

Good stuff to know, with a reasonable explanation why this might be a problem in all brands if they are mostly using a common imager. More is not always better. Shoot, even the 720P camera still looks better than the 1080P camera in the 1 lux chart.

Reactions:

Jon Dillabaugh

Pro Focus LLC

I actually just had a different experience with a Hikvision 4MP Bullet. I was actually pretty amazed at how well it performed in a low light setting recently. It was able to see way more details than my naked eye. The scene was very dark to my naked eye, but the camera was able to see quite a bit of detail in color mode, shortly before it switched to night mode. I was wondering when it was going to make the switch to night mode. Now, that said, I didn't have another camera in that same scene to compare it against 720p or 1080p, but it wasn't a poor performer for sure. I am curious now and feel compelled to do a shootout tonight to compare.

Reactions:

(1)

John Robinson

This topic came up in an article here a few years ago here where IPVM did a comparison of 2, 3 and 5 mpxl cameras, both daylight and low light. I remember making the comment that 2mpxl looked like the best bang for the buck in terms of overall image quality in both environments. I tried to search for the thread but couldn't find it.

Reactions:

Jon Dillabaugh

In reply to John Robinson

Pro Focus LLC

I believe this is the test you are thinking of. It is dangerous to blindly assume higher megapixel cameras will perform worse than lower megapixel units though. From model to model, you will see great variances, regardless of resolution. Some of it will be directly correlated to sensor size. Others will be software and configuration based.

Reactions:

(2)

John Robinson

In reply to Jon Dillabaugh

That's similar but this is the one I was talking about.

https://ipvm.com/reports/ip-camera-line-count-test

Here's my comment:

John Robinson

| 3 years ago

Interesting, in that it would appear that for most purposes 1080p is the optimal bang for the buck, performance wise. Nearly double that of 720p and only slightly less than 5mpxl. Also has the best low light performance.

Reactions:

Undisclosed Manufacturer #1

I remember testing Arecont's binning a couple years ago and thought that did a very good job of low light performance, though we really haven't implemented it in the field.

Is binning proprietary to Arecont. If not, why don't more manufacturers offer it as an option in night mode?

Reactions:

John Honovich

In reply to Undisclosed Manufacturer #1

IPVM

1, the problem is that Arecont (across the board) is comparatively poor in low light (see our most recent Arecont tests). However, binning does help. Not only did it moderately improve low light (for comparatively terrible to poor), it reduces bandwidth which is a big concern at night for most cameras.

Reactions:

Undisclosed Manufacturer #1

Well that sounds like all the more reason for another manufacturer to also offer binning as an option. But does Arecont own the patent on it that prevents others from offering it? Or have others just not thought it that important or useful a feature?

Reactions:

John Honovich

In reply to Undisclosed Manufacturer #1

IPVM

Binning goes far beyond / before Arecont or any IP camera. Additionally, I have not seen Arecont even claim a patent on a specific binning implementation.

As for others, I do not know. The main downside, from a marketing standpoint, is that you are admitting that you won't get full resolution benefits at night anyway, which is clearly technically true but not something many probably want to admit from a general marketing standpoint.

Reactions:

Undisclosed Manufacturer #1

I see your point, but if they test it and it does work, I don't see why it couldn't be an a camera menu option to use or not use. That way customers can see for themselves if it works and use it at their discretion.

Reactions:

Rick Puetter

My company, Pixon Imaging, has a technology that we are trying to push that uses non-destructive reads to create multiple images with different levels of binning from a single exposure. Then a single image is created with high spatial resolution for the bright parts of the scene and lower spatial resolution (but higher sensitivity) for the dimmer parts of the scene. So it would be like taking a simultaneous 4Mp and 720p picture of the scene and using the best parts of each picture wherever you need it in an automatic and spatially-adaptive way. This way you always get the best picture. We have a patent on this and we're looking for manufacturing partners. Anyone interested?

Reactions:

Undisclosed Manufacturer #1

Rick, I assume you have probably approached most of the big names already, for example Axis, Bosch, Panasonic, Sony, Hikvision, Dahua, etc. If that is correct, what has been some of the challenges gaining adoption by those big names?

Reactions:

Rick Puetter

In reply to Undisclosed Manufacturer #1

We are just starting a serious effort to market (sell, license, etc.) our technology (about a two-month old effort), but we are looking at the cell-phone camera market first. We have approached some of the big names there, but we haven't approached many of the security camera manufacturers. Of course there is some overlap, but in any event, we haven't heard back yet.

Reactions:

(1)

John Honovich

IPVM

Rick, what is the rough cost premium of your sensor vs typical sensors used in surveillance cameras? How much of a barrier is that likely to be?

Reactions:

Undisclosed #2

IPVMU Certified

Sounds like off-sensor processing.

Reactions:

Rick Puetter

The cost is essentially the one-time design cost. The per unit cost, once designed, is essentially zero.

Reactions:

Rick Puetter

The multiple readouts with different binning is done on chip. There is then modest off-chip processing comparable to what people already do with sensor isp's.

Reactions:

Undisclosed #2

IPVMU Certified

Let me try for 0 for 2:

If it's on the sensor then I'm guessing you're talking about CCD's?

Because I've heard that binning is stupid hard on smaller pitch pixels in CMOS because of the per pixel circuitry.

Reactions:

Rick Puetter

In reply to Undisclosed #2

Sorry if I didn't pick up on what you were asking.

Yes, this could be done on CCDs, and that's how we first contemplated doing this. But this can also be done on CMOS sensors, and I'm betting that this is where it will be used first. And yes, you are correct, binning is more natural on CCD devices. Nonetheless, while more difficult, CMOS charge-domain binning has been done since the early 80's in research labs, and charge transfer on CMOS front-ends to provide global shutters capability currently is done in a number of main-stream CMOS sensors. So the technology to do in CMOS is proven and available.

Reactions:

Undisclosed #2

IPVMU Certified

No, you picked up fine, my comment was only self-deprecating.

So then is your design something physically different than global shutter CMOS chips have today, or are you saying that that is your requirement?

Reactions:

Rick Puetter

In reply to Undisclosed #2

Laughing. Thanks!

No. Global shutter technology, such as in the Sony Pregius sensors, does charge shifting before converting the charge signal to an output voltage. It is the charge shifting capability that allows binning of charges, and collection of charges from multiple photo-collection sites, onto a single readout gate that provides the capability needed for my technology. So the fact that CMOS charge binning has been done in University labs for decades, and the fact that modern, mainstream manufacturers, e.g., Sony, do charge transfer in CMOS readout front-ends, that demonstrates that my techniques are a viable solution for the poor low-light performance of mega-pixel sensors. Essentially, the multiple exposures use the mega-pixel sensor as designed as a high spatial resolution sensor for bright objects, be then using the same exposure turn the sensor into a lower-resolution sensor with bigger, more sensitive pixels in a second readout of the scene. The two (or more) resulting readouts of the same exposure then can be combined (not a complex procedure) to form a single image with greatly enhanced low-light sensitivity. Further, unlike multiple exposures with different exposure times, since we use only one exposure, there are no motion artifacts between the images.

Reactions:

Jon Dillabaugh

Pro Focus LLC

In reply to Jon Dillabaugh

Rick,

Great info! I do have a question though. What happens in the darker parts of the scene in relation to pixel density? What if I needed 50ppf to identify a given object in my scene, but those pixels fell below the light level threshold and ended up gettin binned. Would I lose my PPF in that scenario and jeapordize my identification? Are there issues with motion blur or any other noise issues?

Reactions:

Rick Puetter

•Jul 26, 2016

Hi Jon,

Well pixel density does go down, but only enough so that you can see something. The philosophy is that it is better to see something at low spatial resolution than nothing at all at high spatial resolution.

Reactions:

Undisclosed #2

•Jul 26, 2016

IPVMU Certified

In reply to Undisclosed #2

Essentially, the multiple exposures use the mega-pixel sensor as designed as a high spatial resolution sensor for bright objects, be then using the same exposure turn the sensor into a lower-resolution sensor with bigger, more sensitive pixels in a second readout of the scene. The two (or more) resulting readouts of the same exposure then can be combined (not a complex procedure) to form a single image with greatly enhanced low-light sensitivity.

I'm confused are you saying that you are or are not able to use someone's silicon as it exists today to implement your innovation?

Is there a prototype?

Reactions:

Rick Puetter

•Jul 26, 2016

We have not been able to find a currently available CMOS prototype. So unfortunately, no, we haven't been able to test this directly on a CMOS device. It is easy to test, however, with CCDs and multiple exposures, since there are lots of CCDs that exist that have binning capability. Not surprisingly, you get exactly what you expect theoretically. This is not rocket science. (I am a rocket scientist, BTW. Well, an astrophysicist.) This is very simple physics.

Reactions:

Undisclosed Integrator #3

•Feb 12, 2017

If a 4mp camera is turned down to 2mp will you get the same night vision performance as a factory 2mp camera or do you need to actually buy a 2mp camera instead, for this example I would say I am using HIK

Reactions:

John Honovich

•Feb 12, 2017

IPVM

In reply to Undisclosed Integrator #3

Generally speaking, reducing resolution streamed of a camera will not significantly improve low light performance. It might stay the same but use less bandwidth which is an advantage and some cameras using binning which can be marginally better.

However, it is hard to get around the limitations of pixel sizes. The Omnivision 4MP is a 1/3" imager so too are most 2MP cameras today, which means the pixels on the Omnivision 4MP are going to be a lot smaller which adds to the low light problem.

Reactions:

(1)

Rick Puetter

•Feb 15, 2017

Hi John,

But there are some important qualifications. If the binning is done on-chip and in the charge domain, then this is exactly the same as having larger, but a smaller number of, pixels. And you correctly recognize this as a powerful solution to the low-light sensitivity problem. It also, of course, reduces band-width since the number of independent pixels are reduced. So in the charge-domain binning scenario, if m x m = N pixels are binned, and you are working in the read-noise dominated regime (usually the case), then the signal goes up by N without any increase in the noise (read-noise in this case). If pixel binning is done in the digital domain, as for example with an on-chip signal processor, then the signal goes up by N, but the noise goes up by SQRT(N)=m, so the signal-to-noise ratio improves by N/SQRT(N) = SQRT(N) = m. This is not as good as binning in the charge domain, but is still a significant improvement.

Reactions:

(1)

John Honovich

•Feb 15, 2017

IPVM