SCALABLE OBJECT RECOGNITION USING SUPPORT VECTOR MACHINES

Automatic recognition of objects in images now typically relies on robust local image features. For scalable search through a large database, image features are quantized using a scalable vocabulary tree (SVT) which forms a large visual dictionary. In this project, we design support vector machine (SVM) classifiers for tree histograms calculated from SVT quantization. We explore several practical ker- nels that naturally capture the statistics of image features. A baseline Naive Bayes classifier for tree histograms is also created for compar- ison. After Naive Bayes or SVM classification, we further perform a geometric verification step to avoid false positive matches, using either affine or scale consistency check. The Naive Bayes and SVM classifiers and the geometric verification algorithms are tested on two real image databases with challenging image distortions.