Image classification using the Bag-of-Features model. Training images are represented using clustered and quantized Speeded-Up Robust Features (SURF). Classification using K-Nearest Neighbor.

Accuracy on a subset of Caltech101 (with 10 classes) is 76.67%. Tested with GNU Octave (4.0.2).

Read the publication here: An Improved Bag-of-Features Approach for Object Recognition from Natural Images

• Download image dataset: https://www.vision.caltech.edu/Image_Datasets/Caltech101/
• Edit main.m and change the dataset_root variable.
• Run main to train and test the system. Displays the confusion matrix and average accuracy.
• Run test to test with individual files.

• OpenSURF version 1c includes its own license.
• Rest of the code is under GNU GPL v3. Read license.txt for details.