Robust object recognition using a cascade of geometric consistency filters

Bag-of-words is a popular and successful approach to performing object recognition. Its performance is limited by not considering relative geometry information. This limitation is particularly stark when there is significant image noise. We propose a “bag-of-phrases” model which extends bag-of-words by enforcing geometric consistency through application of a “geometric grammar” in a filter cascade. Experimental results on a computer generated dataset show increased robustness to clutter and noise as demonstrated by more than two orders of magnitude reduction in false positives compared with bag-of-words.