spatial_envelope

Classify images from

http://people.csail.mit.edu/torralba/code/spatialenvelope/spatial_envelope_256x256_static_8outdoorcategories.zip

using Bag of Words and SIFT

under /report/report.docx

you can find performance measurements using ROC AUC

BOW [Bag of Words]

• Choose 2 categories from the above dataset

• Calculate dense-sift for all picrutes

• Vector quantization - Calculate K-Means for all extracted features. Choose at least K=100 means.

• For each image, calculate a histogram of its features of the k'th cluster.

• Train a linear SVM, where each picture is represented by its histogram, and each picture is labeled by its class.

• For the test phase:

• Calculate SIFT for each image.

• For each feature, find its nearest neighbour.

• For each image, calculate its histogram.

• Classify using now trained SVM.