SplitNet

This is the repository for SplitNet paper.

Prerequisites

0. Caffe

Installation

$ cd caffe
$ "Adjust Makefile.config"
$ make -j

The caffe contained in this repository is modified from caffe-1.0, including

• CornerCropLayer, for "image50" and "image56" setting
• LocalLayer, for "SplitNet-conv52" setting
• linear lr_policy
• extract_txt tool, for saving features to txt file

Running Experiments

There are 5 file in each setting dir:

• train_test.prototxt: The network definition for training. You need modify /path/to/your/traininglmdb and /path/to/your/testinglmdb before using.
• solver.prototxt: The solver definition for training.
• extract.prototxt: The prototxt for extracting feature. You need modify /path/to/your/evaluationlmdb before using.
• train.sh: The script for training models.
• extract.sh: The script for extracting features.

You can train by simply running

$ cd basic (or basic2x or image50 or image56 or SplitNet-conv22 or SplitNet-conv52)
$ ./train.sh

Once the training is completed, you can generate features by simply running

$ cd basic (or basic2x or image50 or image56 or SplitNet-conv22 or SplitNet-conv52)
$ ./extract.sh