Cameras for best AI performance
Ben suggested that for optimal AI I look for 8 MP, P-Iris, and autofocus (as well as careful attention to placement and lighting). He advised me not to worry too much about low-light color capability in comparison since IR B&W does so well.
8MP/P-Iris/autofocus cameras seem rather few and far between, and my highly experienced preferred installer has questioned how important the P-Iris feature really is.
(As an aside, he is also biased against Hikvision, which does have such cameras at a reasonable price, due to disturbing human rights allegations, e.g. Uyghur-specific face detection for the Chinese government's program to control ethnic minorities. He's also concerned about the phone-home problem with the Chinese-government affiliated companies such as Hikvision, though that we can easily solve that I think with simple network segregation.)
I'm curious how important P-Iris and autofocus are to SecuritySpy's AI performance as a practical matter. I'm strongly leaning toward following Ben's advice based on his experience, but it does complicate camera selection.
Thanks, Bill
Comments
My understanding is that P-Iris has primarily developed to increase image quality in changing lighting conditions such as direct sun glare. I haven’t much experience with this hardware myself.
Personally I have an 18 camera install and have made a few recent upgrades. My install is predominantly Hikvision with a few Dahua units.
For the best AI results I’ve stuck with the following main principles, although I agree 8MP is the way to go, auto focus for me only became relevant with verifocal models, it all depends if you can achieve sufficient detailed coverage of your scene with a fixed lens.
Typically I mount fixed 4MM lenses around 4.5M heigh or 2.8MM lenses depending on the scene, these are above head height but low enough to get a detailed view of subjects. Positioning is also so important like not too close to a wall to reduce bounce back from IR. Proper use of image setting profiles to ensure that the ideal settings for day and night are switched when needed. Things like WDR and exposure will differ between the two. These cameras are not plug in and forget if you want decent results. Adjust each to their own scene it’s trial and error but worth the work.
For example many cameras come set to 1/12 or 1/25 exposure which looks great when static but blurs movement at night, especially when trying to pause footage.
I also have PTZs by dahua and wide lens cameras high but only to serve as an overview.
I see many people Installing 5MP Turret cameras at roof or soffit height. That’ll give a great birdseye view but won’t be sufficient for identification. Checkout the D-O-R-I specs of whatever you go with. Make sure you’re happy with the identification metrics. I’ve had great success with the latest Hikvision 8MP ColorVu models - the latest also contain IR they call it Hybrid Light so you can select what works for your scene. Don’t chase megapixels but take into account sensor size! An 8MP camera with a 1/3” sensor will perform poorly compared to a 1/1.8”, the amount of light available to a sensor is key even when using IR.
As for the Hik and Dahua controversy that’s a personal moral call but from a safety prospective I’ve analysed via wireshark and noticed nothing untoward. Calls to WAN are mainly observed when P2P has been initialised and this is something I wouldn’t recommend. Good network segmentation should be completed regardless of brand choice, in combination with keeping firmware up-to date and good password housekeeping should be more than sufficient for most users to mitigate any security concerns. Unfortunately the price to performance ratio is really hard to beat. The likes of axis, although may be ideal but at 4-5x the cost per unit I can’t personally justify it.
i use the AI recording on all cameras daily and regularly review footage. It’s best in class with all the cameras I’ve thrown at it but with an 8MP decent sensor, correctly positioned camera it really shines. If you need any model numbers or want any stills or samples from what I’ve got let me know. I know hardware selection can be a mindfield.
Basically, the clearer the image, the better the AI is going to perform. Resolution is important and 2 MP should be considered absolute minimum (with 4 MP and above being preferable), but a really clear sharp 4 MP image is always going to be better than a grainy blurry 8 MP image, so features like low-light sensitivity and optical quality are very important.
We recommend auto-iris lenses where possible (of which P-iris is a particular variant), because this allows the camera to use smaller apertures in bright conditions during the day, which can significantly sharpen the image.
@JC_5674 makes some good points about camera mounting angle and minimum exposure time, these will both make a big difference to the accuracy of AI classification. Note that shortening the minimum exposure time will result in a noisier image in low light, so there is a tradeoff here. Installing off-camera IR lighting will significantly improve the image at night, allowing for short exposure times without too much noise, so this is definitely something to consider if the on-camera IR lighting is too weak (which it often is).
Customers worry about Chinese-brand cameras for two main reasons:
Thanks JC_5674 and Ben!
"I’ve had great success with the latest Hikvision 8MP ColorVu models" That's exactly what I was thinking, perhaps this: https://www.hikvision.com/en/products/IP-Products/Network-Cameras/DeepinView-Series/ids-2cd7587g0-xzhs-y-/. Not inexpensive, but as you say very reasonable for the advanced optics and features. The promised ability to image well in color even in extremely low light is intriguing, but I'm not convinced it'll allow better AI than good old fashioned black and white mode with excellent IR illumination.
Ben, that's a very clear explanation of the value of the more advanced optics. This house gets tremendous sunlight, inside and out, at times so the automatic iris does sound valuable, and P-Iris sounds most capable by far of maximizing image sharpness. I've seen firsthand the very poor results in bright light with the existing, very basic analog CCTV cameras.
My personal thoughts on the big Chinese manufacturers. Privacy is a concern with all cameras, as Ben says, and easily assured with network segregation, so that's not a reason to shun Hikvision and Dahua. Ethnic face recognition and the like are extremely disturbing to me, but as a practical matter my paltry purchases certainly won't affect the fortunes of these huge companies. And I'm pretty sure all the surveillance companies, from all countries, are enabling governments and other entities to do things that would make my skin crawl if I only knew about them. Like all technology, advanced surveillance tech can be used for good or for ill. And if it can be it will. Besides, just as SecuritySpy implements its excellent AI in software, I imagine anyone with enough resources could deploy things like ethnic face recognition with any brand cameras. These are personal judgments. I certainly have full respect for those who feel otherwise and prefer to buy from other brands.
I'm now trying to choose specific cameras for optimal AI per Ben's suggestions (8MP/P-iris/autofocus). Options seem limited, but Hikvision offers these:
https://www.hikvision.com/us-en/products/professional-ip-and-analog-cameras/intelligent-series/deepinview-cameras/ids-2cd7586g0-izhsy/
https://www.hikvision.com/us-en/products/professional-ip-and-analog-cameras/intelligent-series/deepinview-cameras/ids-2cd7587g0-xzhsy/
The second one (7587) promises great color vision in extremely low light conditions, but I'm not convinced that will help due to motion blur and other compromises compared to good old fashioned IR with good illumination. The positive is that B&H Photo offers the 7587 at a rather amazing discount: https://www.bhphotovideo.com/c/product/1749332-REG/hikvision_ids_2cd7587g0_xzhsy_2_8_12mm_colorvu_ids_2cd7587g0_xzhsy_8mp_outdoor.html. (As an aside I'm not clear how B&H can offer such discounts, but they do appear to be a legit authorized Hikvision dealer).
A possible negative of the 7587 is that the aperture is F1.0 to F1.2, whereas the 7586 model's aperture is F1.2 to F2.5. My question is whether the smaller aperture is actually an advantage for AI purposes in that it allow greater depth of field and, if I understand correctly, potentially reduced lens aberration, one of the very purposes of P-iris technology. Or perhaps we can safely assume the Hikvision engineers have sufficiently optimized the 7587 P-iris performance even with a lens that only closes to F1.2?
If so I think the 7587 should work well since you can apparently set it for black and white, relying only on IR illumination, in low light if that works better than the Hikvision low light color vision.
Thanks, Bill
Also...any other ideas for 8MP/P-iris/autofocus cameras would be much appreciated.
Hi @nagar, both models you link to are high quality and I'm sure will give you excellent results.
Yes, the ability for the camera to set a small aperture is a good thing for image quality: as you say, it provides a greater depth of field, and it will also provide generally sharper images during the day. However note that at night, the camera will have to widen its aperture, probably to the widest value, so a smaller possible aperture is only an advantage during daylight hours.
1.0-1.2 is a surprisingly narrow aperture range; this may not provide too much benefit. On the other hand, the spec sheet of the 7587 does list better low-light performance than the 7586. So it's a bit difficult to choose between these two specific models - perhaps you can let price and availability be the deciding factors.
Thanks very much, Ben.
As a followup I just talked with a seemingly very knowledgeable surveillance camera specialist at B&H Photo. They do sell the 7586 model, which is based on Hikvision's older DarkFighter technology and lacks their ColorVu technology for color vision in near darkness. I was surprised that this model is more expensive than the 7587 with ColorVu ($680 compared to $600).
He wasn't surprised. He said the DarkFighter low light technology is actually superior to ColorVu for maximizing AI analytics. If he's right this would confirm my instinct that Hikvision had to make some serious compromises to get color vision in nearly lightless conditions. The very narrow aperture range did seem concerning. In addition to such optical compromises he said that ColorVu has a heavy duty software processing layer that makes its own compromises to squeeze out color info in the dark.
He said seeing color in darkness is very "cool," and useful for some specific purposes (e.g. police wanting to see the color of clothing or a car). But it's by no means better in general than good old fashioned IR-illuminated black and white tech.
I just came across this comparison of DarkVision and ColorVu, which agrees with B&H's take:
https://www.a1securitycameras.com/blog/colorvu-vs-darkfighter/
Thanks for this interesting information, this does make sense. There are always compromises that have to be made with the design of cameras and optical systems, and if you optimise for one feature (e.g. colour in low light) you may have to compromise on another feature (e.g. detail). Personally I would prefer detail over colour, but others may have good reasons for making the opposite choice. In terms of AI performance, both colour and detail are equally important, so it's difficult to say which would give better results for this specifically.
Thanks, Ben. Regarding color vision and AI, back in June you wrote:
"Colour helps a bit with detection accuracy, but not very much. Sharpness and resolution are much more important than colour, so this should be your priority when choosing a camera and lighting. If the camera produces colour images, but they are grainy or they have a lot of motion blur (because long exposures need to be used), then this will give poor results. I would be surprised if any camera were able to produce good colour images at night at near total darkness - this sounds like an impossibility. Personally, I would go for an 8 MP camera that has good low-light performance but that is also sensitive to IR, and I would install some supplemental IR lighting to boost the quality of night images."
But here you wrote: "In terms of AI performance, both colour and detail are equally important, so it's difficult to say which would give better results for this specifically."
I'm curious: has your view changed on the importance of low light color vision to AI detection? As more companies introduce color vision tech for near darkness I'd think this will become a more important question over time.
In my message above, I should have made it clearer that I was speaking in the context of these two specific cameras, where I'm suggesting there is probably a tradeoff between a little more detail/sharpness on the one hand, and more comprehensive colour rendition on the other hand. In this specific case of these two cameras, it's hard to predict which will give the better AI results.
More generally, I stand by my original statement from June. If Camera 1 produces accurate colours at night, but has a significantly blurry or noisy image, and Camera 2 switches to black & white at night, but provides sharp, noise-free images, then Camera 2 would be the clear winner.
Thanks, Ben!