I'm user the Caffe implementation of GoogleNet available from https://github.com/BVLC/caffe/tree/master/models/bvlc_googlenet in a detection and localization process. I've had great success using this network for transfer learning on my own fine-grained classification task, but I've had a much harder time using the out-of-the-box implementation. Using a slightly tweaked version of deploy.prototxt with batch sized reduced from 10 to 1, I'm predicting the classes for the images contained in this directory.
{
'bikini1': {
'445': 0.44777653,
'459': 0.42523962,
'543': 0.0014282113,
'638': 0.083084606,
'639': 0.03712019},
'bikini2': {
'445': 0.34988937,
'459': 0.56838924,
'638': 0.053508826,
'639': 0.0262988,
'775': 0.00064279145},
'cheeseburger1': {
'923': 0.00077056035,
'924': 0.00074912253,
'931': 0.0040806201,
'933': 0.98862922,
'964': 0.0012325585},
'cheeseburger2': {
'923': 5.947132e-05,
'928': 0.00013152717,
'931': 0.00084521802,
'933': 0.99827468,
'934': 0.00013743658},
'ipod1': {
'487': 0.0010906343,
'605': 0.99825007,
'632': 9.7395357e-05,
'662': 0.00010778452,
'754': 0.00011993563},
'ipod2': {
'487': 1.3061179e-05,
'592': 2.9665862e-05,
'605': 0.99993002,
'681': 9.4235074e-06,
'761': 3.5998178e-06}
}
These images are taken directly from the ImageNet training set. What's interesting is that the network assigns the same label to each pair with a high degree of confidence, but these labels seem totally at odds with the labels provided by ImageNet. The network seems to be working, but maybe I'm referencing the wrong class-label mapping? Any help from all you Caffe/ImageNet specialists would be greatly appreciated!
