This is the CAFFE implementation of "Successive Embedding and Classification Loss for Aerial Image Classification".
-
Download [MSTAR dataset]. Organize the subset of the dataset following Table I of the paper;
-
Randomly partition the training set into real training set (with name "train90_sqrt_mag_aug.h5") and validation set (with name "train10_sqrt_mag_aug.h5"). Name the test set "test_sqrt_mag_aug.h5".
-
Run grid_siamese.sh (note that this will automatically run batches of experiments. Change hyperparameter if you only want to run one or some experiments with specific hyperparameters.)
-
Note that the prototxt of contrastive loss on feature layer can be found at "siamese_feature.prototxt".
- Run grid.sh and will grid learning rate for network trained with center loss on classifier layer.
The use of this software is RESTRICTED to non-commercial research and educational purposes.