OpenCV SaaS - would you use it?

asked 2015-01-23 07:32:31 -0500

cryptid gravatar image

Hello devs,

I've been playing with OpenCV lately and I did notice that a lot of time and resources are spent on things like classifier training, detection, etc. This is especially true when doing calculations on low performance devices. I have some unutilized high performance machines (100+ cpu threads, 40gb memory, etc) that I want to put to a good use, so I was wondering if anyone would be interested in using a service that would allow you to move your OpenCV calculations to 'the cloud'? Kind of like Heroku for OpenCV!

Basic idea is that you would do the calculations remotely either via a webpage or using an API. This would allow you to do things like cascade training in matter of hours instead days/weeks. If this were to gain traction we could expand the idea to cover other things like object detection with custom classifiers and much more, all done somewhere in the cloud without you having to do the coding manually and on the local machine.

Let me know what you think about the idea. Also, if anyone out there is interested in joining forces in such a project, feel free to reach out

edit retag flag offensive close merge delete

Comments

1

+1 like said before I think this can help out people alot! What I would do is also supply feedback on the training parameters and resulting models, some sort of evaluation pdf that you receive. This could give pointers on how to improve training of your model.

StevenPuttemans gravatar imageStevenPuttemans ( 2015-01-23 08:22:04 -0500 )edit

What is the advantage of using your platform instead of amazon?

FooBar gravatar imageFooBar ( 2015-01-24 08:42:07 -0500 )edit
1

Basically you wouldn't need to build your environment manually. For example, with AWS, if you wanted to run traincascade you would need to setup an instance, configure and install OpenCV and only then run traincascade. I could provide all that ready, you would just need to specify your sample.vec. Plus AWS high performance machines are 32 cores max, mine are 170 ;) there are many more advantages, but this is just a start.

What do you think?

cryptid gravatar imagecryptid ( 2015-01-24 10:12:01 -0500 )edit

Sounds interesting enough to do some calculations :) I think you need some numbers to see if the idea can succeed and how much your service can cost. How long does it take to install OpenCV an AWS? (If you have a nice build script?) What could stop Amazon from also providing machines with OpenCV? I've never really trained classifier, so how do they scale with the number of cores? As Steven said, some evaluation tools are important. You could maybe add some of the standard evaluation datasets and analyse the just trained classifiers. You'd then even have some comparable time analysis. I'm one of the authors of the RGBD-Dataset for 3dScanning (http://vision.in.tum.de/data/datasets...) and I think an option to evaluate is really important for many (academic) users.

FooBar gravatar imageFooBar ( 2015-01-25 03:25:11 -0500 )edit