An effective bag-of-visual-word scheme for object recognition

Bag-of-visual-word (BOW) model for object recognition has attracted much attention in recent years. The ambiguity of visual words is a key issue to limit its performance. This paper presents an effective BOW scheme by introducing spatial weights, which is dependent on the saliency map. Different from conventional visual attention regions based on segmentation, this saliency map is obtained from selected visual words after one-vs-all SVM (Support Vector Machine). This segmentation-free saliency map emphasizes the contributions of visual words belonging to foreground, effectively reducing the redundancy and ambiguity of conventional BOW models, and marking the localizations of objects as well.