BagOfVisualWords

Image categorization using Bag of visual words approach.

Uses opencv python-contrib SIFT for feature extraction and scikit-learn's SVM.

This is python implementation of Bag of visual words model, which again is based on the paper by Csurka et al[1].

Please refer here, for details regarding implementation of the model.

Things you need to install before running:

• pickle
• scikit-learn
• opencv-contrib (it contains sift implementation)
• numpy
• matplotlib

Project architecture:

  -BagOfVisualWords/
        |- images/
                  |-train/
                          |-category 1
                          |-category 2
                          |-etc
                  |-test/
                          |-category 1
                          |-category 2
                          |-etc
                  |-kmeans/
                          |-category 1
                          |-category 2
                          |-etc
        |- bag_of_words.py
        |- read_files.py
        |- plot_data.py
        |- all other python and sav files.

Things to remember before running:

1. Change the path to images in the read_files.py file, to point to the directory containing test, train, kmeans images.

2. Change the path in the KMeans_clustering.py to point to the directory containing kmeans images.

Usage:

python bag_of_words.py

Additional Information:

• I have also included the cluster center files, which were obtained using KMeans_clustering from the sift features of 1000 images.

• They are named as bov_pickle_Numberofclustercentres.py. (Number of clusters being 200, 400, 600, 800). Feel free to use them.

• All the data required for testing, training and also for vocabulary building has been collected from mostly Caltech101 and few images from Caltech256.

Results:

Number of categories	Accuracy	No of clusters used
3	83%	600
3	85%	800
4	76%	600
4	76%	800
5	81%	600
5	80%	800
6	67%	600
6	67%	600

• Please check the results folder for the confusion matrices obtained. ex: confusion matrix of 5 categories with 600 clusters.

.

Documentation:

• Each python file contains the information about what each function does, what arguements each fucntion takes and how they can be tweaked.

References:

• (1)[https://www.cs.cmu.edu/~efros/courses/LBMV07/Papers/csurka-eccv-04.pdf].

Feel free to open an Issue if you face any problem. :)