Recognition of many small objects on the whole image

[Дмитрий Маренков]

Hi guys! How to recognize small objects on whole image? 

I have two ideas:

1. Use the find objects algoritms  + CNN (CNN receives a set of rectangles)

2. Trying to learning network on whole image.

What's better? Perhaps someone has had experience in this area.

I would be glad of any help!

[Bartosz Ludwiczuk]

Hi,

I have some experience about such problem.

1. You mean getting the object proposition from algorithm and then use it for classification. From my test it will not work good, mainly because of small object which are not detected by Object Proposition algorithm. In fact,  Andrej Karpathy analyse of detection from ImageNet 2014,  show that the models have very big problems with small objects. And they are using Object Proposition algorithm.

2. I think this is much better, to classify responses for whole image. Then it need to be done in multi-scale. And you need a lot of training data. I am thinking exacly about sth like OverFeat paper. You can also try the modification of OverFeat:  https://github.com/Russell91/ReInspect

Or you can create your own method. 

[Дмитрий Маренков]

Thanks, Bartosz! 

Here is another interesting work - http://arxiv.org/abs/1406.5726

пятница, 6 ноября 2015 г., 10:21:56 UTC+3 пользователь Дмитрий Маренков написал: