All geek questions in one place

Does convolutional neural networks have localization capabilities on images?

As far as I know, CNN relies on sliding window methods and can only indicate if any pattern is present or not in the specified bounding boxes. It's true?

Is it possible to achieve localization using CNN without any help from such methods?

+6
source share
2 answers

This is an open problem in image recognition. In addition to sliding windows, existing approaches include predicting the location of an object in an image as CNN output, predicting boundaries (class pixels that belong to the image border or not), etc. See, for example, this article and the links in it.

Also note that when using CNN using max-pooling, you can determine the positions of function detectors that contributed to the recognition of objects, and use them to determine the possible location area of ​​the object.

+8
source

There are several recent methods for locating objects in CNN. See This Article http://cnnlocalization.csail.mit.edu/Zhou_Learning_Deep_Features_CVPR_2016_paper.pdf

It uses a layer called Global Average Pooling (GAP), and without additional work, CNN can localize the object that it recognizes.

Also check out this really good blog post: https://alexisbcook.imtqy.com/2017/global-average-pooling-layers-for-object-localization/

+2
source

Source: https://habr.com/ru/post/981623/

More articles:


All Articles