23. R. Sukthankar: Boosting Distance Metrics for Improved Recognition

There has been significant recent interest in learning appropriate distance metrics from data for improved matching, retrieval or recognition. We present a novel method that generates a weighted Hamming space by boosting many thresholded projections of the data.
We will present some theoretical results showing why this is a good idea and experiments that show significant benefits of applying this method in classification and retrieval. Interestingly, one can view many learned descriptors as special cases of distance metric learning.

This is joint work with Liu Yang and Rong Jin.