Wei Zheng, Changhu Wang, and Lixin Chen
Automatic object detection for an arbitrary class is an important but very challenging problem, due to the countless kinds of objects in the world and the large amount of labeling work for each object. In this work, we target at solving the problem of automatic object detection for an arbitrary class without the laborious human effort. Motivated by the explosive growth of Web images and the phenomenal success of search techniques, we develop an unsupervised object detection framework by automatically training the object detector on the top returns of certain image search engine queried by the name of the object class. In order to automatically isolate the objects from the Web images for training, only clipart images with simple background are used, which keep most of the shape information of the objects. A two-stage shape-based clustering algorithm is proposed to mine typical shapes of the object, in which the inner-class variance of object shapes is considered and undesired images are filtered out. In order to reduce the gap between clipart images and real-world images, we introduce an efficient algorithm to synthesize the real-world images from clipart images, and only shape feature is used in the detector training part. Finally, the synthetic images could be used to train object detectors by an off-the-shelf discriminative algorithm, e.g., boosting and SVM. Extensive experiments show the effectiveness of the proposed framework on objects with simple and representative shapes, and the proposed framework could be considered as a good beginning of solving this challenging problem.
Publisher IEEE Internation Conference on Multimedia & Expo (ICME)