Adaptive Performance-Based Classifier Combination for Generic Object Recognition

It is well-established in the pattern recognition community that the performance of classifiers can be greatly improved by combining the outputs of multiple classifiers. In this paper, we introduce the concept of adaptive performance-based classifier combination, i.e., the weighting of classifiers based on their estimated recognition performance, to generic object recognition. Using an expectationmaximization (EM) algorithm previously applied to image segmentation, we learn the characteristics of individual generic object recognition classifiers. Using the ETH-80 data sets we demonstrate that by incorporating these performance estimates in a Bayesian classifier combination, the recognition rate of the combined classifications improves substantially over feature combination as well as simple and confidence-based voting. The EM algorithm has no tunable parameters and does not require a pre-classified training set during the learning stage. We conclude that adaptive performancebased classifier combination is a valuable and versatile tool to improve the performance of generic object recognition systems.