Amazon Rekognition Custom Labels - LEGO brick classifier in AWS

[@MirceaAdam]

Thank you for creating this video, it was exactly what I was looking for. Great content and explanations. Now I know how to proceed with my own project.

[@tsunghsiwu]

thank you for the demo and explanation. worth of my time in less than 20 minutes to have a nutshell of AWS Rekognition!

[@abhiramivs_ponytales]

Very crisp and nice explanation Mike! Thank you for the video.

[@SohailImtiaz]

Thank you and bravo ! Very simple And to the point explanation of key benefits and pricing.

[@mikegchambers]

The link to the course was broken, sorry. It’s fixed now. link.mls-c01.com

[@cloudjourneyman395]

Awesome. Can't wait to see the robot in action. :)

[@mikegchambers]

Thanks! Me too!

[@sosaucily]

Very cool, definitely the best video I've seen so far! For a hobbiest, who wants to leave an infrequently used service up and running all the time paying this high hourly rate, but who does want custom labeling and the auto-training features, what service would you recommend?

[@angelubilla5586]

Amazing video Mike! Thank you very much!

[@mauriciobertolucci]

Thank you very much for sharing this knowledge in such a simple way Mike. I am precisely in the position of wanting to use a cell phone APP, which should be available 24x7. Isn't there a way to run the trained model on our own infrastructure, like an EC2?

[@EnglishwithNate]

great video! It makes me wonder what use case this could have, like, why not have it identify the next lego in the instruction? could it identify a picture of a picture of a lego, how about multiple legos? So many questions!

[@sundarmeha]

Great video and explanation than any other AWS Rekognition videos out there. One question, the command line response was just the label and the confidence score. How did you pull up the correct image out of the returned custom labels? what if multiple custom labels are returned? I have a similar use case and this would be greatly useful

[@nadiagill]

Hi Mike - Happy New Year! Bhups(LinkedIn)

[@vizlifestudios]

Fantastic! This helped answer a few questions I had about the service before experimenting myself.

[@jbla7986]

Who have you created all those images? And what script did you use?

[@hippomaster100]

this video is awesome wtf. How did you generate the automatic 3d images on EC2??

[@kakshay7]

Please share code for the above project. Looks very interesting.

[@vitor.brangioni]

Hi, Mike. Nice video, tks! I have a question... When is better use Tensorflow instead Amazon Rekognition Custom Labels? Wichs is advantages to use Tensorflow instead it?

[@mikegchambers]

Great question. So… Custom Labels works well if you don’t have a heap of ML experience, and you want to make inference within an AWS centric system. If you have specific needs, and want to be able to run the model anywhere, including in and out of AWS, then a Tensorflow model works well, and you can use SageMaker for the development, training and hosting of you want.

[@shanalive]

Where can I get information on how to do the mobile front end please please please

[@luciamaniscalco5346]

It is possible to count legos? For example in a complex image with differents legos i can obtein how many bricks for every type?

[@mikegchambers]

Well this certainly is possible. The system starts to get more complex of course. There is an amazing iPhone app that does just that: it’s actually called the Brickit App if you search in the App Store. I’d like to build a more complex model that could do this level of object detection. One day. :)

[@talha_anwar]

learnt a lot, cant we move these architectures to some where else like to local, ec2 once is trained

[@mikegchambers]

Hello. Glad you enjoyed the video. It would be nice if you could but unfortunately you cant move these models from the service. If you want to host the model locally or on EC2 then you can use something like SageMaker to create your own model.

[@mikegchambers]

Previously you asked about the labels… you can set them as you like, I used official LEGO part numbers, because, well, because why not. :)

[@talha_anwar]

Thanks a lot for quick response. I really like your idea of robotics with AI for legos

[@ozgurgeke8692]

Hi Mike, that is perfect. Thank you so much for that. I do have a question though, can you also measure the volume of those lego objects perhaps with LIDAR with your phone? If so how? Thank you.

[@mikegchambers]

I’ve played around with LIDAR scanning quite a bit, and particularly with the latest iPhone and iPad. Unfortunately the accuracy just isn’t there for small objects like this. Depending on what accuracy you need you’re probably better off taking images of the object on a background with a known pattern (such as a checker board etc) and then processing the image to approximate its size.

[@chriskiepfer450]

Hi Mike, what a great video. I am setting a path to build a sorting machine. This will help tremendously. Could you share which 3d program you used and maybe the script that you used to generate the images? Thanks Chris

[@felipecamposclarke]

Thanks Mike, great video. very clear. I only have one question, which 3D software or library did you use to generate the images?

[@mikegchambers]

I’m using Blender most of the time. That’s my new setup anyhoo.

[@markuskinast2008]

​@@mikegchambers Thank you for this video. Is there a library at Blender or where did you get the library with all the Lego-bricks/models from ? When the Lego Bricks are in, what Programm/Script did u use to generate the Background, turn Bricks and finally generate the picture ?

[@viktorbranco8362]

Hey. Do you think AWS Rekognition is better than Google's free Teachable Machine? Judging from this video it seems definitely better but it's also true it's not a free service.

[@mikegchambers]

Different services have different pros and cons. :)

[@yvesclaes8554]

Really nice video, Thanks. I'm interested on how you generate the dataset. Is there somewhere I can get the scripts you used?

[@mikegchambers]

Yeah I worked on it for a while to tidy it up. Maybe I can release it one day. :)

[@kezy2695]

could you share code for that app, which is sending image to the aws?

[@mikegchambers]

I go into basically the same code here:

[@mikegchambers]

kzbin.info/www/bejne/f2HMf2isfqZ-prs

[@mikegchambers]

Let me know if that works out, otherwise I’ll see if I can post another video next year.

[@camilofigueroa1289]

I have a code with a basic php example (with API secret and key). That is working with AWS default tags and objects. I have created a custom model with my custom tags, it is trained too. I was read and seeing many KZbin videos and AWS documentation, but none can say me how to connect my trained model with my PHP (API key and secret). The "use" button from my model show me a Python code, but i dont need the Python code. I have the ARN, I have create a new API, but that road is uncertain. My custom model joined with my php (which take pictures) is the last step for my project. Please help me to use my model in the PHP program.

[@mikegchambers]

This should be what you’re after: docs.aws.amazon.com/aws-sdk-php/v3/api/api-rekognition-2016-06-27.html#detectcustomlabels

[@darylteo2219]

So what is the alternative if you need a live inference model with more than 250 custom labels and 24/7 uptime? Thank you!

[@mikegchambers]

In the AWS world you can train up an image classifier or object detection model using SageMaker built in algorithms. That’s the easiest way. Other than that you can build custom models in your ML framework of choice.

[@darylteo2219]

@@mikegchambers Seems like SageMaker only allows 250 classifiers as well (probably where the 250 limit for Rekognition comes from) so it does not seem fit for purpose here as you say. But thank you, we'll probably have to roll out our own OpenCV solution as you say. 🙏 thank you for taking the time to reply!

[@mikegchambers]

Interesting. I’ve not seen that limitation. Can you point me to where it says that for SageMaker? I’d like to dig a bit deeper.

[@darylteo2219]

@@mikegchambers docs.aws.amazon.com/general/latest/gr/sagemaker.html here it says 250 images? did I misunderstand that wrongly?

[@darylteo2219]

@@mikegchambers Aha no I am a dummy, SageMaker Images refer to container images, not picture images.

[@stanislavhinov4225]

Do you have a tutorial on the mobile frontend?

[@mikegchambers]

Not sure why I didn’t get notified about this comment… but as it happens yes, it was in my last live stream video. :)

[@michaeljuliano8839]

Span is not the past tense of spin. Spun is.

[@mikegchambers]

Indeed!