Feature Extraction with Caffe

Image feature extraction with Caffe
Including some examples classify the Oxford-IIIT Pet Dataset using scikit-learn

Dependencies

Caffe, Python 2, NumPy, scikit-learn, matplotlib

Installation

Download the CaffeNet modelfile to caffe/models/bvlc_reference_caffenet/ and other dependent files by this script

➜ ./data/ilsvrc12/get_ilsvrc_aux.sh

Modify the deploy.prototxt file as following

➜ cp models/bvlc_reference_caffenet/deploy.prototxt  
      models/bvlc_reference_caffenet/deploy_feature.prototxt
➜ gedit models/bvlc_reference_caffenet/deploy_feature.prototxt

# line 152
layer {
  name: "fc6"
  type: "InnerProduct"
  bottom: "pool5"
  # top: "fc6"    # comment out
  top: "fc6wi"    # add
  inner_product_param {
    num_output: 4096
  }
}
layer {
  name: "relu6"
  type: "ReLU"
  # bottom: "fc6" # comment out
  bottom: "fc6wi" # add
  top: "fc6"
}

Usage

Make sure that the path to caffe_root (-c), images (-p) are correctly designated.
You should prepare a '.npy' file (-i) contains image filenames in numpy.array.
If you use a GPU, set the device ID (-g).

➜ python feature_extract.py -c caffe/ -p images/ -i image_filenames.npy -g 0

Examples

Contour of grid search (SVM)
Classification of Oxford-IIIT Pet Dataset using SVM

> accuracy = sklearn.metrics.accuracy_score(test_labels, predicts)
> print(accuracy)
0.829838709677  

> report = sklearn.metrics.classification_report(test_labels, predicts, target_names)
> print(report)
                            precision    recall  f1-score   support

          american bulldog       0.66      0.80      0.72        50
 american pit bull terrier       0.66      0.58      0.62        50
              basset hound       0.78      0.80      0.79        50
                    beagle       0.71      0.60      0.65        50
                     boxer       0.71      0.78      0.74        50
                 chihuahua       0.89      0.78      0.83        50
    english cocker spaniel       0.81      0.84      0.82        50
            english setter       0.85      0.78      0.81        50
        german shorthaired       0.95      0.82      0.88        50
            great pyrenees       0.84      0.84      0.84        50
                  havanese       0.88      0.84      0.86        50
             japanese chin       0.93      0.86      0.90        50
                  keeshond       0.98      0.98      0.98        50
                leonberger       0.88      0.92      0.90        50
        miniature pinscher       0.87      0.82      0.85        50
              newfoundland       0.85      0.92      0.88        50
                pomeranian       0.95      0.78      0.86        50
                       pug       0.96      0.98      0.97        50
             saint bernard       0.81      0.86      0.83        50
                   samoyed       0.84      0.94      0.89        50
          scottish terrier       0.90      0.90      0.90        49
                 shiba inu       0.79      0.88      0.83        50
staffordshire bull terrier       0.56      0.66      0.61        41
           wheaten terrier       0.79      0.84      0.82        50
         yorkshire terrier       0.96      0.92      0.94        50

               avg / total       0.83      0.83      0.83      1240

Confusion matrix
Keeshond (the best class in F-score)
Staffordshire bull terrier (the worst class)
American pit bull terrier (sometimes predicted as staffordshire bull terrier)

livst / feature-extraction-with-caffe Goto Github PK