Curious about Dahua's new AI ipcam lineup

[thendawg]

Not sure if any of this information is known, but Ive been unable to find it via Dahua. If Im just missing a pdf or something feel free to link me

What type of method is being used to analyze for object detection, post process after a motion event or frame by frame (realtime) analyzation?

Any idea what type of inference theyre using? (Like RCNN, YOLO, etc)

Platform? (tensorflow, spark, etc)

Any info on the hw being used for this?

Has anyone done any (reasonably) long term testing for accuracy rate?

Im just trying to compare to my current ML person detection project (which has been quite successful at the expense of hw) and get some idea of what it takes to scale (accurate) ML inference based detection to cheaper hw.

 

[john-ipvm]

We've not seen Dahua share any specific lower level technical details and Dahua offers a variety of cameras. That said, for those lower-cost Dahua 'AI' cameras (say $500 or less), I doubt they are using anything like YOLO or Tensorflow, etc., mainly because the processing required and the cost to do so on-board is restrictive (at least right now).

From what we have seen of video surveillance manufacturer marketing generally, 'AI' is often used broadly to include any type of machine learning algorithm, not specifically single shot object detectors like YOLO.

We tested the high-end Hikvision one (DeepInView, retail US ~$900) recently and it was using a Myriad 2, worked fairly well for Chinese video analytics but still inferior to Avigilon, e.g. We have not tested any of the Dahua one's yet as we try to wait until they are officially sold in the US to avoid any objections about unauthorized / China only products being unrepresentative.

 

[thendawg]

Thanks for the info! I can also understand these companies being tight lipped as well, so it seems reverse engineering is really the only way to figure things out. Sounds like I may need to pick one up lol. Interesting about the Hik - Ive actually been researching the Myriad 2. I've also been looking into the Cyclone 10 as it seems to have had some great success in ML applications, but the dev board is expensive, and I literally have 0 FPGA knowledge - Ive been learning a bit, but getting your feet wet is more like diving in the deep end and hoping you can swim hah. Myriad seems to have a nice C based SDK, so I could definitely get on my feet quicker that way, I just need to learn more about its limitations.

Oh and about the marketing term "AI" I TOTALLY agree Certainly a broad definition that manufacturers are likely using to their advantage.

 

[john-ipvm]

For the Myriad 2 and the newer Myriad X, Intel makes Neural Compute Sticks that make it easy to try out. They are ~$100 for the USB devices, reference Intel® Movidius™ Neural Compute Stick

 

[thendawg]

For the Myriad 2 and the newer Myriad X, Intel makes Neural Compute Sticks that make it easy to try out. They are ~$100 for the USB devices, reference Intel® Movidius™ Neural Compute Stick

Thanks! This post has decided to make me give the Movidius platform a try, ended up picking up the neural compute stick 2 for $88 on Amazon, should be here Wed. Not sure if it would be appropriate somewhere on this forum or not, but would love to startup a post with my progress.