Who's Using Intel Openvino For Video Surveillance Analytics?

Last month, I asked Who's Using YOLO For Video Surveillance Analytics? and we had a good discussion.

Now, I am asking about Intel OpenVino. Here's their marketing video for those who want a quick overview:

To answer my own question, we hear a lot of manufacturers are working on using OpenVino. I am curious to hear about how companies are using it and what benefits or issues they are seeing.

We are working on an upcoming OpenVino post and have already begun our own testing of OpenVino via the Movidius compute stick.

Agree
Disagree
Informative: 3
Unhelpful
Funny

We've had the Intel folks reach out to us to test Movidius with our platform (I'm a co-founder of Deep Science -- we do audio and video threat detection with deep nets). While very cool for a low power USB stick, it doesn't have the memory or the horsepower to run the larger networks in real-time, which is a necessity for us to seriously do false positive suppression. 

I haven't seen any reason why OpenVino would offer an advantage for adoption. Like NVIDIA, Intel plays the silly game of showing OpenVino inference time improvement in comparison to CPU-only inference, rather than other specialized hardware such as GPUs. 

Agree: 2
Disagree
Informative: 2
Unhelpful
Funny

Sean, good feedback!

The stick is using the Myriad 2, not the newer, more powerful Myriad X. I think that some manufacturers are going to use those for at least some applications. For example, Avigilon has committed to Intel / Movidius for their H5 upcoming camera generation.

Sean, are you doing the processing / inference in the cloud then? AWS, Google Cloud, etc.? What about an on-site appliance? Do those have sufficient power for your needs?

Agree
Disagree
Informative
Unhelpful
Funny

John,

We started with the naive hope of doing our heavy processing in the cloud, but transitioned to a mini-PC with GPU on-site appliance once the realities of bandwidth in retail hit us. We can manage 4 camera streams in addition to audio classification simultaneously for a box under $1k (installation and maintenance is done via a central monitoring station partner). 

I'm hopeful for a near future where we have more processing options for analytics. Intel is definitely on the right track and I look forward to the day where their devices have the memory to run the more accurate networks. Perhaps then they'd be willing to do the apples-to-apples comparison against NVIDIA when it comes to inference speed, rather than CPUs. 

Agree
Disagree
Informative
Unhelpful
Funny

John,

In AxxonSoft we do use OpenVINO Toolkit. Our reasons:

1) It is not so easy to convince a partner to buy expensive NVidia Tesla cards those are intended to be installed into server platforms. At the same time NVidia GeForce cards are not fully suited to be used in 24/7 mode in a server. As the result parthers search for an alternative solution - not so expensive but stable.

2) It's better to use Tesla cards via NVidia TensorRT package. But there is no a Windows version of TensorRT and Windows is the main choice of our customers.

3) OpenVINO is the hardware abstraction layer for us. It may use CPU, Intel GPU (Intel HD Graphics), Movidius, FPGA cards. Currently Intel GPU and Movidius (based on Myriad 2) are not able to compete with a good NVidia card, but support of OpenVINO makes us ready for the next generation of devices from Intel those (I hope) are coming.

4) On CPU OpenVINO is really much more efficient in comparison with Caffe
which we currently use for inference on NVidia GPU. And it is able to fully utilize multi processor servers. So in some projects OpenVINO may be a good alternative to a graphics card.

 

We are working on an upcoming OpenVino post and have already begun our own testing of OpenVino via the Movidius compute stick.

I am looking forward for the results. Especially if you managed to get a Myriad X based device.

Agree
Disagree
Informative: 1
Unhelpful
Funny

Igor, thanks for the details. That is helpful.

Intel just released the NCS 2 (using the MyriadX) yesterday. We bought one and it arrived today so we are hoping to release a test next week on it.

Agree
Disagree
Informative: 2
Unhelpful
Funny

Igor, thanks again.

Some feedback via NVIDIA:

NVIDIA says they added Windows support with Tensor RT v5 that was released last month.

Regarding GeForce cards not being suited:

Recently we launched Tesla T4 that significantly improved perf/power/$ over last year generation. Fully backward and forward compatible across generations. Geforce gaming cards have their own value proposition (basically sold 1 at a time to consumers), but not recommended in an enterprise setting where guaranteed long-term operation, RMA, special support, guarantee of supply, etc. might be necessary.

Also, we published the Intel Neural Compute Stick 2 / Movidius AI Test on Wednesday.

Agree
Disagree
Informative: 1
Unhelpful
Funny

Thank you, John. That is interesting.

Agree
Disagree
Informative
Unhelpful
Funny

Looks like Camio does - https://www.camio.com/intel/

 

Agree: 1
Disagree
Informative
Unhelpful
Funny

Just returned from the Honeywell Connect conference for dealers.  OpenVino was discussed in detail for the Maxpro VMS and NVR products; Intel even was a featured speaker at the opening general session for the commercial side.

Agree
Disagree
Informative: 2
Unhelpful
Funny

Who is OEMing these new NVRs?  Dahua or Hikvision?

Agree
Disagree
Informative
Unhelpful
Funny

Neither. They are being made inside of the US by a company based in Fremont CA. The new NVRS are Intel server based. Previous Maxpro NVRS were made by Dell. The embedded/performance stuff is a different story.

Agree: 1
Disagree
Informative
Unhelpful
Funny