Caffe-TCNN. Using filter banks in convolutional neural networks for texture classification

This is a basic example of the Texture CNN (T-CNN) network developed in "andrearczyk2016using". We provide an implementation of the T-CNN-3 (3 convolution layers) fine-tuned or from scratch on the kth-tips-2b database.

Getting started

Install caffe (caffe-installation) and go to the caffe directory. (Alternatively, you can copy the folowing folders to your already existing caffe folder: caffe-TCNN/data/kth-tips-2b/ caffe-TCNN/examples/kth-tips/2b and do the same following tasks).

Create a folder into which you will import the T-CNN-3 model pretrained on ImageNet:

mkdir ./models/tcnn

Download and untar the caffemodel from this link: tcnn3.caffemodel and place it in the created folder ({caffe-root}/models/tcnn).

Prepare the data (download the kth-tips-2b database, untar and convert to jpg images):

cd ./data/kth-tips-2b
. ./prepare_data.sh

Create the lmdbs and mean files:

cd ../../
. ./examples/kth-tips-2b/create_kth.sh
. ./examples/kth-tips-2b/make_kth_mean.sh

Now 4 sets of lmdbs (in {caffe-root}/examples/kth-tips-2b) and means (in {caffe-root}/data/kth-tips-2b) should be created. The first fold (kth_test1_lmdb, kth_train1_lmdb and kth_mean1.binaryproto) is linked to be trained on and tested. It’s now ready to train.

Evaluation

You can fine-tune from T-CNN-3 pre-trained on ImageNet:

./build/tools/caffe train -solver ./examples/kth-tips-2b/solver_tcnn3.prototxt -weights ./models/tcnn/tcnn3.caffemodel -gpu 0

or

Train from scratch:

./build/tools/caffe train -solver ./examples/kth-tips-2b/solver_tcnn3_scratch.prototxt -gpu 0

To test another fold, link another set of lmdbs and mean and run the same training.

Citation

If you use the provided code in your research, please cite

@article{andrearczyk2016using,
  title={Using Filter Banks in Convolutional Neural Networks for Texture Classification},
  author={Andrearczyk, Vincent and Whelan, Paul F},
  journal={arXiv preprint arXiv:1601.02919},
  year={2016}
}