On the speed and accuracy of object recognition when using imperfect grouping

This paper analyzes the improvements that can be gained in object recognition through the use of simple, imperfect grouping techniques. We consider, in particular, the pose clustering method of object recognition. Simple grouping techniques are described that determine pairs of points that are connected in the image edge map. We show that such grouping techniques can considerably improve both the speed and accuracy of object recognition. Experiments are described that demonstrate the improvements in performance