Show HN: megaAI – a powerful, small, AI-capable 4K camera for Pi

Show HN: megaAI – a powerful, small, AI-capable 4K camera for Pi

megaAI with a US quarter for scale. Actual size: 43mm by 30mm

megaAI with a US quarter for scale. True measurement: 43 mm by 30 mm

The megaAI is a turn-key pc imaginative and prescient and synthetic intelligence decision that mixes and harnesses the 4 TOPS (Trillion Operations Per 2nd) of AI processing vitality with a difficult 4K, 60 FPS digicam for human/object monitoring in a minute, low-energy, gear. It’s splendid for hobbyists and researchers and is prepared for dispute integration by OEMs. It’s additionally appropriate with our DepthAI ecosystem, and is because of this truth insanely easy to exercise.

{Hardware} Sides

  • Extreme Compute Potential: 4 TOPS (Trillion Operations Per 2nd)
  • Onboard Digicam: 12 MP 4056×3040 Auto-Focal degree (8cm-∞) 81-Deg. Horizontal Space of Look (HFOV)
  • RAW Output Specs: 4K @ 60FPS or 12MP @ 30FPS with USB3.1 Gen2 Host
  • {Hardware} Picture Encoding: 12 MP JPEG encoding
  • {Hardware} Video Encoding: 4K H.265 encoding at 30 FPS
  • Low Vitality Design: 2.5 W max
  • Light-weight: 2 oz.
  • Compact Measurement: 45 mm x 30 mm
  • Versatile Make Utilization: constructed for ease of prototyping AND integration into new merchandise
  • Speedily Connectivity: USB3 Type-C 3.1 Gen1 (5 Gbps)
  • Beginning Present: MIT Licensed hardware, instrument, and documentation

Unbelievably Easy Object Detection

The megaAI takes beforehand powerful pc imaginative and prescient duties esteem valid-time object detection and monitoring and makes them as straightforward as plugging in a USB cable and working a Python script. Appropriate clone the DepthAI git repository and dawdle python to go looking a dwell demonstration of MobileNetSSD being dawdle in your host system. It is probably you will nicely most almost definitely most almost definitely even doc dwell 4K, 30 FPS video of all of the items the digicam sees.

Object Monitoring & Detection

Object localization is the aptitude to know what an object is and the place it’s miles within the bodily world. The megaAI is able to kill this at 30 frames per second on a Raspberry Pi, with out together with any load on the Pi.

Absolutely the most sensible-to-dawdle accumulate on megaAI is what we train as our examine case: Object Detection on 20 courses (PASCAL VOC 2012):

  • Specific particular person: particular person
  • Animal: rooster, cat, cow, canine, horse, sheep
  • Automobile: aeroplane, bicycle, boat, bus, automobile, bike, put collectively
  • Indoor: bottle, chair, eating desk, potted plant, couch, TV/visible present unit

Which implies, whilst you be part of megaAI to a number and degree the digicam at these objects, you’ll look a bounding field with a label drawn spherical every object.

Pre-trained Detection & Monitoring

Want different neural community units? There are lots of pre-trained units that may work applicable type away, many are accessible today from Intel without cost. Swap the brand new mannequin in to the built-in Python script and progress! You’re able to move.

Some examples of frigid megaAI-compatible devices:

  • COVID-19 disguise/no-hide detection
  • Age recognition
  • Feelings recognition
  • Face detection
  • Facial Landmark (e.g., corners of eyes, mouth, chin, and hundreds others.)
  • Typical object detection (20-class)
  • Pedestrian detection
  • Automobile detection

Personalized Detection & Monitoring

At the same time as you esteem to take care of one thing customized, likelihood is excessive you will nicely be able to put collectively your non-public units consistent with accessible/public datasets after which train OpenVINO to deploy them to DepthAI. Uncover our documentation piece on creating custom tracked objects.

True-time H.264 & H.265 Encoding

Now not best is megaAI a quickly and energy-efficient methodology to position in energy inventory or customized recurrent neural networks, it’s additionally a valid-time video encoder that compresses 4K H.265 video at 30 frames per second.

Extreme Body Prices, Low Vitality Consumption

Since megaAI does so mighty processing by itself (together with optionally compressing video), the load on every USB and the CPU is diminished vastly, permitting Luxonis to assemble dramatically elevated body prices, bringing low-energy {hardware} acceleration to any venture:

True time object detection with OpenVINO and Movidius

Pi 3B+, CPU BestPi 3B+, NCS2, OpenVINOPi 3B+, DepthAI, OpenVINO
MobileNetSSD (show on)5.88 FPS8.31 FPS25.5 FPS
MobileNetSSD (show off)6.08 FPS8.37 FPS25.5 FPS

(Raspberry Pi/NCS2 recordsdata courtesy of the righteous of us over at PyImageSearch)

Some Concepts For How To Use megaAI

  • Computerized 4K Filming – Beginning recording when best explicit issues or actions are seen. E.g., that splendid Nationwide Geograhic shot (of your housecat)
  • Agriculture/meals processing – Check for applicable apples and sinful apples, win show of apples harmed by pests for stat capabilities
  • Efficiently being/safetyAre of us carrying masks?? Detect unsafe stipulations
  • Mining – Together with mining for Legos – going through a pile of Legos to go looking out the ‘gold’ one
  • Manufacturing – Excessive-tail valid-time defect detection, look for explicit defects, and keep the system improve over time

Movidius Myriad X, Unleashed

DepthAI, megaAI’s instrument, permits the train of the elephantine vitality of the Myriad X. This grew to become the one mission of the DepthAI venture on the beginning set and we accomplished this through customized implementation in any respect layers of the stack ({hardware}, firmware, and instrument). After reasonably quite a lot of iteration and and collaboration what resulted grew to become an environment friendly and straightforward-to-exercise system which takes elephantine simply applicable factor in regards to the 4 Trillion Operations Per 2nd (TOPS) imaginative and prescient processing potential of the Myriad X.

{Hardware} BlockA type of Myriad X OptionsLuxonis DepthAI
Neural Compute EngineCertainCertain
SHAVE CoresCertainCertain
Movement EstimationInaccessibleCertain
Edge DetectionInaccessibleCertain
Harris FilteringInaccessibleCertain
MIPI ISP PipelineInaccessibleCertain
JPEG EncodingInaccessibleCertain
H.264 and H.265 EncodingInaccessibleCertain

Comparability to Another Fleshy AI/CV Reply

MetricmegaAI w/ Raspberry Pi*Intel NUC AI Equipment
Picture Determination12 MP (4056×3040)1 MP (1280×720)
Video Determination4K720p
Easy Setup & TypeCertainNo
Ambiance wonderful Data RouteCertainNo
True TimeCertainCertain
Low LatencyCertainCertain
CPU Free for Specific particular person CodeCertainNo
CPU UtilizationCome-zeroExtreme
{Hardware} H.265 StrengthenCertainNo
{Hardware} JPEG StrengthenCertainNo
{Hardware} Function Monitoring StrengthenCertainNo
Vitality6 W (max, together with Pi)50 W (max)
Stamp$169 USD (appropriate through advertising marketing campaign)$879.95 USD

A Raspberry Pi is now not built-in with any megaAI product or pledge supplied appropriate through this advertising marketing campaign.

Beginning Present – MIT Licensed

megaAI is an supply supply venture. So whereas you esteem to favor to originate one thing off of it, enact it!

We’ve supply sourced {hardware} and instrument to allow you to enact so:



Full documentation on the general instrument

And even our documentation is delivery-offer, so whereas you search an error likelihood is excessive you will nicely be able to publish a PR with the repair!

From Zero to Synthetic Intelligence

It’s by no means been simpler to win up and working with machine discovering out, pc imaginative and prescient, and synthetic intelligence.

With megaAI it’s applicable type a handful of steps sooner than you’re up and working. The subsequent video reveals all of the items required with a Raspberry Pi.

For different applications (macOS, Residence home windows, and Linux variants), it’s applicable type as straightforward. How about teaching megaAI to detect customized objects?

That’s straightforward too. We offer free on-line teaching through Google Colab notebooks:

The below tutorials are consistent with MobileNetv2-SSD, which is an object detector which natively runs on DepthAI. A bunch of different object detectors might be educated/supported on Colab and dawdle on DepthAI, so whereas likelihood is excessive you will nicely keep a rely on for a very totally different object detector/community backend, please certainly be contented to publish a GitHub Space! We’re always together with additional (mainly as of this writing, we applicable type purchased some new YOLO variants working).

Easy Object Detector Teaching: Birth in Colab

The academic pocket book Easy_Object_Detection_With_Custom_Data_Demo_Training.ipynb reveals how likelihood is excessive you will nicely be able to quick put collectively an object detector consistent with the Mobilenet SSDv2 community.

After teaching is full, it additionally converts the mannequin to a .blob file that runs on our DepthAI platform and modules. First the mannequin is reworked to a construction usable by OpenVINO known as Intermediate Illustration, or IR. The IR mannequin is then compiled to a .blob file the train of a server we arrange for that motive. (The IR mannequin may even be reworked in the neighborhood to a blob.)

And that’s it, in decrease than a pair of hours likelihood is excessive you will nicely keep an developed proof of thought object detector that may dawdle on megaAI to detect objects of your totally different. Nonetheless don’t protect our uncover for it, abet discovering out for an instance we constructed.

Occasion: COVID-19 Conceal/No-Conceal Teaching: Birth In Colab

The Medical Conceal Detection Demo Teaching.ipynb teaching pocket book reveals an instance of a additional sophisticated object detector. The teaching recordsdata set consists of of us carrying or now not carrying masks for viral safety. There are practically 700 images with roughly 3600 bounding field annotations. The images are sophisticated: they differ so much in scale and composition. Alternatively, the factor detector does reasonably an correct job with this considerably diminutive dataset for such a job. Another time, teaching takes spherical 2 hours. Looking out on which GPU the Colab lottery assigns to the pocket book occasion, teaching 10okay steps can protect 2.5 hours or 1.5 hours. Each methodology, a brief length for such an correct great proof of thought for such a troublesome job.
We then carried out the steps above for changing to blob after which ran it on our DepthAI module.

Risks & Challenges

Manufacturing at scale is mostly a expose, with problems taking place at high-volumes which had been merely now not expert at smaller manufacturing. We will enact our best to nimbly reply to those problems as they come up and abet to the agenda, whereas handing over megaAI on the ordinary of great we at Luxonis would question as prospects.

Now we keep already produced a primary dawdle of the megaAI boards, so our supply chain and manufacturing activity has been examined and verified efficiently.
Pointless to assert, delays are that likelihood is excessive you will nicely be able to mediate with any manufacturing activity, and we’re able to debate any concerns with backers through well-liked updates.


The megaAI is a ready-for-assemble product. We delivered all 4 well-liked DepthAI boards with additional elements above and past these promised appropriate through that advertising marketing campaign. Now not best that, however in distinction to another crowdfunded campaigns, we shipped on time!

Logistics: Transport & Achievement

The megaAI boards might be fulfilled within the following batches:

  • Amount in first manufacturing batch: 500
  • Estimated transport date of first manufacturing batch: 30 July, 2020
  • Amount in subsequent manufacturing batches: 100
  • Present agenda of subsequent manufacturing batches: 30 August, 2020

NOTE: Roadrunner tier pledges might be shipped as we speak on the conclusion of the selling marketing campaign by Luxonis (based in USA). The aid of the selling marketing campaign orders might be shipped the train of Crowd Present’s fulfillment carrier. Residence (appropriate through the US) orders ship with out value, whereas world shipments incur a transport surcharge that’s utilized at checkout.

Thanks For Studying All That – Right here’s Some Excessive Secret Information

COMING SOON: A PoE mannequin of megaAI!


  1. This is a single board based around the Myriad X chip, and the CrowdSupply campaign launched on Thursday. I wrote a lot of the copy and created the video. I think it's an awesome product for hobbyists, researchers, and enthusiasts like me. I'm planning on hooking it up to a Pi-powered robot tank body, with the megaAI attached to a tilt/pan servo.

    These boards are interesting to me in part because they give us important parts of AI recognition/tracking without the severe problems that cloud processing brings. It makes bespoke and embedded uses so much easier.

    I think what HN will be most interested in is that this is from a very small team of highly experienced/skilled people, but we've already made big updates to the software and hardware offerings since the successful campaign for DepthAI last year. All the software is MIT-licensed, and several of our boards are fully open designs you can find on our Github (


Please enter your comment!
Please enter your name here