Combined Segmentation and Visual Attention for Object Categorization and Video Semantic Concepts Detection

Recent researches show that the benefits of image segmentation have been exploited in object categorization and recognition approaches. In most of these works, objects are segmented from the background around to increase recognition accuracy. However, it is generally hard to find a segmentation that captures all correct object boundaries in images of real world scene. So some researches begin to choose several segmentations for representing the objects and performing object categorization. In this paper, we take advantage of an efficient graph-based algorithm for image segmentation, and combine a visual attention model to locate the salient and effective segmentations in a real world image. We propose a model which extends the bag-of-features method for modeling the semantic objects. We evaluate our approach on two experiments: multiclass categorization in Caltech 101 datasets and high-level features extraction in video datasets of TRECVID2007. The results show that combining segmentation and visual attention makes our model achieve competitive performance.