From Bag of Categories to Tree of Object Recognition

Abstract

To recognize different category of objects, multiclass categorization problem is often reduced to multiple binary problems. Traditional approaches require training different classifiers for each category. This can be slow and the performance of learned single classifier is poor for limited training samples. We present a multiclass object recognition tree, in which the leaf node and the non-leaf node correspond to one category and a bag of categories, respectively. Each non-leaf node captures the shared features of a bag of categories. Each node also holds a group of classifiers trained by AdaBoost, to discriminate the categories locating at its left and right child node. Recognition is then a process to find a path from the root to a leaf, which represents a unique category. The very promising result on Caltech 101 dataset shows the robustness of the proposed approach.