Amazon Rekognition – Image Detection and Recognition Powered by Deep Learning | Hacker News Reader

Amazon Rekognition – Image Detection and Recognition Powered by Deep Learning

aws.amazon.com

182 pointsdwynings44 comments

Pricing (for first million images / month)

Google Cloud Vision $1.50 per 1,000 images[1]

Amazon Rekognition $1.00 per 1,000 images[2]

Interesting that when Google Cloud Vision first came out it was $5 per 1,000 images (for label detection), but they reduced it on Nov 15, 2016. [3]

[1] https://cloud.google.com/vision/pricing [2] https://aws.amazon.com/rekognition/pricing/ [3] https://cloudplatform.googleblog.com/2016/11/Cloud-Machine-L...

Microsoft's offering is $1.50/1000 calls as well. https://www.microsoft.com/cognitive-services/en-us/pricing

If you guys want to see an example of this in action I made a stupid little demo. You can tweet @AWSCloudNinja a selfie and it will replace your face with a ninja mask.

I'm waiting to post the code until the new boto models are out but it's all in python and lambda functions.

(I work for AWS)

Can it put fruit hats on kitties?

No but I code a thing that would put hats on people. https://mrmonksy.com/p/ProjectFunnyHats#.WD9gybIrLRY (It was a project for a class)

There are a lot of services like this (Google Cloud Vision, IBM Watson, CloudSight, Clarifai, Imagga) but I'm yet to see one that lets me provide my own dataset. That's what I'm most interested in.

So far the only solution I know of is to take Google's Inception net and train/host it myself somewhere.

I know Watson at least lets you do this: https://www.ibm.com/watson/developercloud/doc/visual-recogni...

It does but it's extremely cumbersome if you have densely tagged images, as you need to send each image once per tag. Plus I believe there's a limit to the size of a ZIP.

Not sure what you mean here - With Watson Visual Recognition, you can train a set of custom classifiers from your example images.

Then in a single API call, you can upload a batch or single image and get a score for each image for each class in all your classifiers. There are limits to the size of the ZIP files you can upload, that's true.

Maybe you were referring to the _demo_ of custom training, here: https://visual-recognition-demo.mybluemix.net/train ? The demo only classifies using one classifier at a time, but the API can do multiple at once.

https://developer-preview.clarifai.com/guide/train#add-image...

Try this - Clarifai is run my Matthew Zeiler, who's imagenet results really kickstarted the whole deep learning craze.

Clarifai's web interface for training with your own images is great, although a little slow — but that could be my computer.

I've used their standard/general model for classifying most of the images for the search on my site, and would love to use custom models for various side project ideas I have, but the price is way too much for me ($0.46 per concept, per month).

Hey, Jason from Clarifai here. Reach out to us. I'd be interested in hearing more about your ideas, and how we might be able to support them. Fun fact: we're willing to give volume discounts based on number of concepts, inputs (indexed image search results), or predictions (tagged images).

Thanks for getting in touch, Jason. That's good to hear! I'll be sure to contact you when I've sifted through my ideas.

Metamind used to let you do that. Clarifai lets you host models as well AFAIK.

https://developer-preview.clarifai.com/guide/train#train

Imagga can help you with this. Http://imagga.com/solutions/custom-categorization.html

Have seen this, haven't tried (and I have no affiliation): https://www.trycleo.com/

It really depends what you're interested in accomplishing. If you want to DIY a self-hosted, custom classification solution then TensorFlow, Caffe, Torch7, etc are your best bet. ClarifAI, Watson, Google Cloud, Imagga will classify images into defined buckets through their hosted API. Other cognitive services, like http://CloudSight.ai/, focus on providing a deeper understanding to the image (fine-grained object recognition or scene captioning) using a massive library (currently over 350M+ images) in order to accomplish the task, though we can also utilize custom libraries.

Disclaimer: I'm the co-founder of CloudSight. We, of course, would also be happy to discuss a solution with you as well. You can contact us at: contact@cloudsightapi.com

Worst AWS product name after Elastic BeanStalk.

It's just named after the product they bought:

https://www.bloomberg.com/news/articles/2016-04-05/amazon-sa...

The ReKognition API

And so? They didn't name RedShift (great name!) after the company they acquired.

I don't care either way--I was just pointing out that they didn't choose the name

https://console.aws.amazon.com/rekognition gives me a "Please (re)authenticate or return to https://console.aws.amazon.com" error.

Anyone else getting this or is it not open yet?

Yeah. Same. Similar thing happened to me when lambda came out. I'm guessing they are still rolling it out to everyone.

Anybody know whether it is kosher to use Rekognition (or comparable Google/Azure/Clarify solutions) as a source of "supervised" data to bootstrap your own in-house ML models? I'm pretty sure the TOS would frown upon that, but it would beat manually tagging your own datasets, plus its hard/impossible for them to enforce.

EDIT: in retrospect, the pricing means its probably not economically viable.

Are there other similar services out there or is this completely new?

It looks really awesome.

I know this is silly to ask of a cloud company but some kind of offline mode would be really useful for robotics and for the security example. Maybe it could be limited like you request n specific things you want to recognize and it stores the relevant data locally.

This seems comparable to Microsoft Azure Cognitive Services: https://azure.microsoft.com/en-us/services/cognitive-service...

Google's Vision API: https://cloud.google.com/vision/

Might as well use a pretrained Imagenet or Inception for an offline mode. Of course it would be harder to use.

Clarifai: https://developer-preview.clarifai.com/quick-start/

Check out Torch7 (http://torch.ch) or Google's new TensorFlow (https://www.tensorflow.org) - both of them are really easy to use (especially TensorFlow) and will allow you to build models with a bit of programming knowledge that run offline quite easily.

"appears to be female - 100%" "does not have a beard - 99.2%" :D

Awesome, I never knew what happened to these guys!

We used to use it to count faces and get some basic demo data from the photos our animated GIF photo booth takes. One day it shut down but we didn't know why until now!

Feel free to give www.kairos.com a try. We have a world class facial detection and recognition service, as well as emotion analysis on still images and video.

(Disclosure CTO of Kairos.com)

Will this help me tag and classify my thousands of personal photos? Once it recognizes a face will it tag it and identify it the next time I call it with the same person on another picture?

Google Photos does that.

Yes.

> You can also use Rekognition to compare faces and to see if a given image contains any one of a number of faces that you have asked it to recognize.

Does anyone know (I don't see in docs, but maybe?) if this has the power to be able to make an image hash for comparison to other images? To be able to do tineye type stuff?

It is based on Orbeus's technology which was acquired by Amazon. Orbeus was found by two Chinese guys in Boston University.

Is there an AWS product that is similar to catchoom.com or slyce.it

Something that recognizes specific items rather than generic terms?

The closest service provided by Amazon to achieve that I'm aware of is Mechanical Turk (https://www.mturk.com/mturk/welcome). It's of course a completely different approach, but depending on the problem it might be a viable solution as well.

This comes rather fast on the heels of Yahoo releasing their NSFW source code

AWS Rekognition's core technology is coming from Rekognition.com, and the company Orbeus behind it was acquired by Amazon in 2015. https://www.google.com/webhp?sourceid=chrome-instant&ion=1&e...

Reply on news.ycombinator.com