Object Detection Using Multiview CCA-Based Graph Spectral Learning

摘要

Recent years have witnessed a surge of interest in semi-supervised learning-based object detection. Object detection is usually accomplished by classification methods. Different from conventional methods, those usually adopt a single feature view or concatenate multiple features into a long feature vector, multiview graph spectral learning can attain simultaneously object classification and weight learning of multiview. However, most existing multiview graph spectral learning (GSL) methods are only concerned with the complementarities between multiple views but not with correlation information. Accurately representing image objects is difficult because there are multiple views simultaneously for an image object. Thus, we offer a GSL method based on multiview canonical correlation analysis (GSL-MCCA). The method adds MCCA regularization term to a graph learning framework. To enable MCCA to reveal the nonlinear correlation information hidden in multiview data, manifold local structure information is incorporated into MCCA. Thus, GSL-MCCA can lead to simultaneous selection of relevant features and learning transformation. Experimental evaluations based on Corel and VOC datasets suggest the effectiveness of GSL-MCCA in object detection.