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.mand change thedataset_rootvariable. - Run
mainto train and test the system. Displays the confusion matrix and average accuracy. - Run
testto test with individual files.
- OpenSURF version 1c includes its own license.
- Rest of the code is under GNU GPL v3. Read
license.txtfor details.