I guess you might have <a href="https://www.engadget.com/intel-realsense-shutdown-211525552.html?guccounter=1&guce_referrer=aHR0cHM6Ly9hcHBsZS5uZXdzLw&guce_referrer_sig=AQAAAF-imOf76303CJMXvN9v9uFtc0knlDeWsr3v8h4QPiwnXeE2KR8rjZqLJwhPNiyJV2gEEdVrdhaxSZd-mvFjeL02sR-LLYXqlfJawhjF_3KAb9yYwqvYDzEq-xEYj70yW5jQ2-nAoF_mpIfM_RvNOkOjz4_q5C3Pw5MuHS-mmNQ_" rel="ugc noopener nofollow" target="_blank" rel="ugc noopener nofollow" target="_blank">read that Intel is Winding down on its RealSense camera business . Given the core processing hardware in those cameras is Movidius chip which is also used in OAK cameras, I was just wondering what will be the continuity of OAK cameras in light of Intel's decision?

Intel Winding Down RealSense

Shaadee

I guess you might have read that Intel is Winding down on its RealSense camera business.

Given the core processing hardware in those cameras is Movidius chip which is also used in OAK cameras, I was just wondering what will be the continuity of OAK cameras in light of Intel's decision?

erik

Hello Shaadee ,
this is the answer from Brandon, our CEO:
Fortunately Luxonis' has your back should you use our tech. We guarantee availability of the OAK-D series through 2028. (And likely well past then, but at a bare minimum through 2028). So very likely our current series will be only used in "Wow, that's legacy!" by the time we do a last-time buy/build (which we will do - to give even more continuity).
link here.
Thanks, Erik

stevefs

Glad to hear this. Very reassuring to one of my customers who I turned on to OAK-D. They have bought 25 OAK-D units. They have some applications they will now likely transition from Intel RealSense to OAK-D.

Intel pulling out of RealSense could give a huge bump to your sales.

I noticed you plan to add support for AprilTags. What about Aruco markers? My customer uses them extensively. Someone posted some code for Aruco tags on GitHub 5 months ago, but I would imagine this was Gen 1 not Gen 2.

Love your products! OAK-D-POE is killer in factories!

stevefs

Given that RealSense is dead will Intel continue to develop new generations of MyriadX?

Brandon

stevefs

Thanks for the kind words. So yes we could likely implement ArUco as well. It's not high on the list right now, but we're in the process of recruiting some junior engineers and this would be a good task.

And on Movidius - it's actually a totally different group than RealSense. And it has always been a bit of weird internal competition between the two. As the RealSense D4 chip (that powers the D415, D435, D455, etc.) is 2 generations older than the Myriad X.

What Intel did was take the Myriad 2, the D4, and made the Myriad X - which is all the parts of the Myriad 2, but with a ton added as well as all the parts of the D4, but combined with all the improved Myriad 2 stuff.

So then this left the RealSense series in a weird spot - where the Myriad X has a dual-core AI, high-resolution/frame-rate ISP, 20+ CV accelerators, a dual-core CPU and the a more-powerful version of the D4 built in. And for not too dissimilar a price-point. So that made for some "awkward hallway conversations" between the RealSense and Movidius teams. ("Wait, your absorbed all our tech? And we absorbed none of your... and now you're selling at the same price as us?")

As Movidius had all the advantages of having built-in dual-core CPU, built-in dual-core AI, 20+ CV accelerators, and the capability for the Movidius chip to be the only chip in a product (and run Python3.9 scripts onboard) - and even if looking at depth-only, it has more than all the capability of the D4 chip.

So if I were to guess, I'd say this is at least one of reasons why Intel would "wind down" RealSense while keeping Movidius going. The current-gen Movidius line is about 2 generations newer than RealSense - and can do more than RealSense in every possible way for the same or lower price-point.

And this is not even to mention KeemBay, which is the next generation, which as Erik mentioned will be in 2022 products from Luxonis.

Oh and for availability of Movidius. So the Movidius team has gotten integrated into some huge things silently, which has resulted in it having gauranteed production for quite a while. Luxonis can gaurantee production into 2028, for example.

Thoughts?

Thanks,
Brandon

erik

Just asked internally about Aruco markers, as I am not sure. We started with gen2 in early 2021, so it could be gen2 as well, but it would only be possible to run detection on host-side (not on the device), as we haven't added support for it yet. WRT new gens of Myriad - yes, Keem Bay should be out in several months, and from my understanding, there are some exciting new features! 🙂 And thanks for the kind words!

stevefs

Brandon

Thanks for your insights!

Keem Bay will be interesting. Sounds like the tflops will go up for the same power consumption.

What things are now run device side? I recall reading somewhere that some OpenCV code now runs device side. What parts of the OpenCV library run device side? Is this run as interpreted Python or C/C++?

Brandon

Great question stevefs,

Here's the list of on-device capabilities that exist now and will exist soon:
https://docs.luxonis.com/en/latest/pages/faq/#what-hardware-accelerated-capabilities-exist-in-depthai-and-or-megaai.

As a note, CPython is now supported, as our edge node and feature tracking, both below, running entirely on-device (no host load):

So OpenCV cannot run device side. Instead we have a pipeline builder where you specify how you want your pipeline to flow (what neural networks, which cameras, the topology, etc. - full node/graph based) and this is loaded to the camera and executed.

So this can be arbitrary combinations of depth processing, neural networks, CV functions, and then Python scripts.

We have a bunch of example pipelines here and here but since it is node/graph based the permutations are literally infinite. We've had customers build things we had no idea would be possible.

We even have community members like Cortic who have made it so that you can make the whole pipeline using visual programming like below (here)

And then here is a good example of a pipeline that uses depth, multiple neural models (palm/dorsal object detection, then hand-pose) and the Python script node to orchestrate the connection of all of this.

So the above is run entirely on DepthAI (he's using the OAK-D-IoT-40 there, but the same pipeline will run on any OAK-D model, including the OAK-SOM series for custom product integration).

Thoughts?

Thanks,
Brandon