Divide 1000 images into 5 unknown classes

[Mạnh Tú Vũ]

Hi all,

I got a problem, I have 1000 images unlabeled. I want to divide those image into 5 classes. No precondition. 

For example: if the machine thinks that the two images A and B are the same class, put it together in one folder. Otherwise, put in a different folder. And finally, I want to have 5 folders contain my 1000 images.

I don't know if Caffe can help me to solve this problem, but if not, please suggest me a paper or link, software, etc if you know. 

Thank you,

[Przemek D]

What you are interested in is called clustering, which is a domain of unsupervised learning, which in general is a rather difficult problem. I think this can be done (depending on scale) using an autoencoder and t-SNE on the encoded features.
But please don't take this answer as an expert's advice (since this is not exactly my area), more like an overview on what topics you could research to find a solution.

[Atena Nguyen]

This is exactly clustering problem like Przemedk D mentioned. One way to solve this problem is using well-known classification network like VGG. First, you calculate the features (final layers in VGG) of each image.

Secondly, using this learning features with another clustering method like the k-mean clustering, etc. (which is available in scikitlearn). 

and if you google the keyword: Caffe + clustering you might find this work Deep Embedded Clustering [1] which use caffe as DL tool. 

Hope it helps. 

[1] https://github.com/piiswrong/dec

Vào 21:21:59 UTC+9 Thứ Tư, ngày 19 tháng 7 năm 2017, Mạnh Tú Vũ đã viết:

[Mạnh Tú Vũ]

I understand, thank you so much :)

[Mạnh Tú Vũ]

Thank you so much. Now I know how to continue my work.