Proper LFW train/test split in Caffe

[Łukasz Kowalewski]

I'm trying to prepare LFW dataset for CNN face recognition network. I've managed to create an lmdb files for both training and test data using the caffe's *convert_imageset* tool. The tool uses txt file as input with the following format describing images and their classes:

    <picture name> <classID>
    01_pictureOfClass1.jpg 1
    02_pictureOfClass1.jpg 1
    01_pictureOfClass2.jpg 2

On the LFW website, they suggest a concrete split for matching pairs for training and testing data, so in this example, it could be something like:

    #Matching pairs
    01_pictureOfClass1.jpg 02_pictureOfClass1.jpg
    #Mismatching pairs
    01_pictureOfClass1.jpg 01_pictureOfClass2.jpg

 1. Is there any way to enforce Caffe to learn certain pairs of matching and mismatching pairs of data? 

 2. If not, how then should I split the whole data of 5549 classes into training and testing datasets?

 3. They are multiple classes in LFW which consist of only one picture. How should I treat the single picture classes in LFW database? Can I even train network to recognize such classes if they are only represent in testing dataset?

 

  [1]: http://vis-www.cs.umass.edu/lfw/ "LFW"