Recognizing an untaught object (image recognition)

asked 2017-01-27 05:36:29 -0500

arvids.p gravatar image

updated 2017-01-27 09:13:20 -0500

I am working with image recognition and classification. My current task is to detect an object that is not in any of the trained classes in order to catch objects that shouldn't be there (view img. 1 where the camera would look at the objects one by one). My current idea is to raise an alarm if none of the trained classes have a high confidence, but this seems like a bad idea. Has anyone dealt with a task like this? Or maybe just share some thoughts on it?

image description image 1

edit retag flag offensive close merge delete

Comments

1

Actually that is exactly what you need. You will always need to limit your classifiers accuracy. Just be sure your classifiers return high scores for the actual object class and not for others (adding them as negatives to your training set)

StevenPuttemans gravatar imageStevenPuttemans ( 2017-01-28 08:26:05 -0500 )edit

Thank you for the response! Have you tried anything like this? Should i try to teach a class with some random objects that shouldn't be there or just use low confidence for other classes as the trigger?

arvids.p gravatar imagearvids.p ( 2017-01-29 02:22:04 -0500 )edit
1

Yep, what we basically do is add classifiers for each known object. If not a single classifier has a confident detection,then your object to be classified will be labeled as background. In a conveyor belt setup you could then use BG/FG segmentation to notice there is actually an object passing by, and label it as unseen object or so.

StevenPuttemans gravatar imageStevenPuttemans ( 2017-01-29 10:02:57 -0500 )edit

What precision can you reach with this method?

arvids.p gravatar imagearvids.p ( 2017-01-30 02:01:07 -0500 )edit
1

How can I tell? Start your implementation and try? Machine learning / computer vision is not an exact sience, you cannot calculate efficiency/accuracy beforehand...

StevenPuttemans gravatar imageStevenPuttemans ( 2017-01-31 06:44:46 -0500 )edit