bag of visual words

A fixed-length feature vector built by clustering local descriptors into a vocabulary and histogramming which cluster each descriptor falls into.

A bag of visual words represents an image as a fixed-length vector by first clustering a large collection of local descriptors into $k$ clusters (the "visual vocabulary"), then, for a given image, building a histogram counting how many of its own descriptors fall into each cluster -- turning a variable number of local features into one fixed-length vector a classical classifier can act on directly.

Examples

Clustering a million SIFT descriptors from a training set into 500 visual words lets any new image be summarized as one fixed 500-length histogram.