Why this picture is classified to Upperbody?

asked 2012-11-15 01:27:34 -0500

fangang190 gravatar image

updated 2012-11-15 12:16:25 -0500

Adi gravatar image

There aren't any people in this image, but it is classified to Upperbody.
We just want to detect is there are any human in this picture.
Can anyone help me with this?
haarcascade_upperbody is what I am using.

image description

edit retag flag offensive close merge delete



Classifiers, like the one you are using, are statistical learners - they fuse a large quantity of information to make up a statistical model of what a human looks like. But as you have discovered, they are not bullet-proof. Usually, their accuracy range is between 20% and 95% - far less than even a child can achieve. And you should also note that they do not use the same features as humans to classify an object, so they can mis-classify an object even when you cannot see the any similarity between the training image and the test one

sammy gravatar imagesammy ( 2012-11-15 13:47:00 -0500 )edit

Thanks for your comments. So is there anyway to improve accuracy in this situation? Or what kind of model rather than classifiers should I use?

fangang190 gravatar imagefangang190 ( 2012-11-18 22:48:19 -0500 )edit

This is a try-and-learn approach. And mainly, developing a specific application in object classification means finding the right classifier + the right features. If I or anyone else could know that without extensive testing, they would build the ultimate robot for themselves :) So, read more papers on the subject, try to implement them, try to understand the good/weak points of them, try to combine them, selecting the best parts of each, and after a while you'll have a much better understanding of what works and what doesn't work.

sammy gravatar imagesammy ( 2012-11-19 04:50:16 -0500 )edit

It sounds like a lot of course work and paper reading. Thanks for your comments.

fangang190 gravatar imagefangang190 ( 2012-11-19 09:20:15 -0500 )edit