Combining different interesting point detectors for object categorization

Many interesting point detectors have been proposed in the literature. It is unclear which detectors are more appropriate and how their performance depends on the task. We propose to use different detectors to gain different cues of images. Then an ensemble of classifications can be obtained, each based on one cue. The use of classification ensemble to categorize new images can lead to improved performance. Detailed experimental analyses on several datasets show that our ensemble approaches are well resistant to the variations in view, lighting, occlusion and the intra-class variations and achieve state-of-the-art performance in categorization.