Author(s)

Sun Shuwan, Jiang Mingyan

Corresponding Author

Mingyan Jiang

ABSTRACT

Exemplar-SVM (E-SVM) was first introduced by Malisiewicz et al, and achieved good performance in the application of object detection and Content-Based Image Retrieval. E-SVM is about training a linear SVM with a single positive sample and many negative samples. All the E-SVMs thus form an ensemble that performs well. Training a good E-SVM requires the negative set to be very large. In this paper we propose to apply E-SVM with augmented features to the area of image detection by softly forcing it to be constructed form existing classifier parts cropped from previously trained classifiers. Our method helps to improve the generalization of the original E-SVM.The proposed method is evaluated on the subset of PASCAL VOC 2007 to detect different orientation of a specific object, and it achieves better performance over E-SVM whilst helps reducing the size of negative set.

KEYWORDS

Exemplar-SVM, Feature Augmentation, calibration.