Development of Image Retrieval System by Considering People Co-occurrence Relations through Relevance Feedback

集体照片通常包含与社会相关的人，如家人和朋友。我们开发了一个从合影集中检索包含特定目标人物的图像的系统，该系统考虑了人的共现关系。从每幅图像中提取代表人的面部外观及其共现关系的人袋(BOP)特征。通过使用BOP特征，可以训练用于将图像分类为两类的分类器，即包含目标人的图像和其他图像。将由分类器确定为包含目标人的图像作为检索结果呈现给用户。通过由用户选择的少量正确检索的图像来重新训练分类器。当从1791张合影中检索32个人的图像时，经过5次反馈迭代后，平均准确率从20%提高到70%。

Group photos often contain socially related people such as family and friends. We developed a system for retrieving images containing a specific target person from a group photo collection by considering people co-occurrence relations. A Bag of People (BoP) feature, which represents both the facial appearances of persons and their co-occurrence relations, is extracted from each image. By using the BoP features, a classifier for classifying images into two classes, images containing the target person and other images, can be trained. The images determined to contain the target person by the classifier are presented to the user as the retrieval results. The classifier is re-trained by a small number of correctly retrieved images which are selected by the user. When retrieving images of 32 persons from 1791 group photos, the mean average precision improved from 20% to 70% after five feedback iterations.