点击次数：

论文类型：会议论文

发表时间：2010-08-23

收录刊物：EI、Scopus

页面范围：44-49

摘要：In many applications, such as object detection, face recognition and visual tracking, it is often desirable to detect the regions of interest (ROIs) in an image. If the sample image is regarded as a bag with its regions being regarded as instances, the problem of detecting ROIs can be viewed as a multi-instance learning (MIL) problem. However, because it is necessary to output the label of each instance, only a few published MIL algorithms can be used to detect ROIs. In this paper, an innovative MIL algorithm is proposed by using Gaussian process. In the proposed algorithm, the class label of each instance is exploited by a latent function with Gaussian process prior over instance space. Experimental results on two object detection problems show that the proposed algorithm is validated and can achieve higher success rate compared with the published MIL algorithms.