Object detection for other specific objects, possible?

[kolleamm]

How would I get started in object detection of specific objects? I know about OpenCV, it can perform face detection, but what about other objects?

I've seen talks on the web of training the program, but I'm not exactly sure how I would do that. I would like to give my software, written in C#, the ability to detect smiles.

There are practically no tutorials for this in C#.

Any help is appreciated.

 

[Simon Bridge]

The first thing you do with any programming project is check to see what others have done before you.
ie. https://www.csie.ntu.edu.tw/~fuh/personal/FaceDetectionandSmileDetection.pdf

 

[Spinnor]

My Google Chrome browser gives me the following warning when I click on the link above,

 

[Mark44]

You can also try this link: http://www.csie.ntu.edu.tw/~fuh/personal/FaceDetectionandSmileDetection.pdf

 

[Fooality]

Just a warning, almost everything in CV is a hack... Valuable things like face recognition have had loads of time thrown at them so they're decent hacks, (and you may have luck with smiles, valuable for marketing) but general object recognition mathematically has no simple solution, its dependant on prior learning.

For instance suppose you see an image of a basketball in the air at a game passing in front of a sign in the stands that's the same color: your brain will recogize the round shape of the ball because you know what a basketball looks like and see its a game, but without that world knowledge no algorithm can. Consequently, the big players in this are the people with huge amounts of data to train AIs about world knowledge. So cloud based solutions from big companies:
http://googleresearch.blogspot.com/2016/03/train-your-own-image-classifier-with.html?m=1
Will probably the way forward, as unsatisfying as it may be.

 

[kolleamm]

The first thing you do with any programming project is check to see what others have done before you.
ie. https://www.csie.ntu.edu.tw/~fuh/personal/FaceDetectionandSmileDetection.pdf

Yeah but those just explain how the algoritm works.

Just a warning, almost everything in CV is a hack... Valuable things like face recognition have had loads of time thrown at them so they're decent hacks, (and you may have luck with smiles, valuable for marketing) but general object recognition mathematically has no simple solution, its dependant on prior learning.

For instance suppose you see an image of a basketball in the air at a game passing in front of a sign in the stands that's the same color: your brain will recogize the round shape of the ball because you know what a basketball looks like and see its a game, but without that world knowledge no algorithm can. Consequently, the big players in this are the people with huge amounts of data to train AIs about world knowledge. So cloud based solutions from big companies:
http://googleresearch.blogspot.com/2016/03/train-your-own-image-classifier-with.html?m=1
Will probably the way forward, as unsatisfying as it may be.

From my research on CV you give the program thousands of images of the object and thousands of images not containing the object. Afterwards it runs a long search lasting days that finds a handful of features that could be used to identify the object. Its a lot of data to provide but finding some code on how to do it might be harder.

I'd love to find something as simple as :
Train_detector (obj imgs path, false positive imgs path)

The result would be a file you can reference
Object_detect (obj_with_these_features.txt)

 

[chiro]

Hey kolleam.

Have you looked at data mining and classification algorithms?

A lot of this stuff involves quite a lot of mathematics in different fields and in different context of application.

 

[newjerseyrunner]

Yes, OpenCV gives you raw image data, from that you can apply whatever algorithm you want. I once used it to keep score in billiards by tracking the balls.