Object recognition on clutter background

Katarina Mele (2003) Object recognition on clutter background. MSc thesis.

Preview

PDF Download (4Mb)

Abstract

Learning and recognition of 3-D objects with appearance-based methods rely solely on visual properties of the objects. Visual appearances of an object are acquired by a sufficient number of images showing the object under different conditions and in different situations including varying angles of poses, changing illumination, different distances of the object from the camera, and cluttered background. In the present work appearance variations are conditioned only by different object poses and cluttered backgrounds. As a learning method the Support Vector Machine (SVM) is used. The classic SVM classifier does not distinguish the background from the object. By changing the view angle the shape of the object silhouette in an image changes. Consequently, the number and the locations of the objects and background vary from view to view. On condition that both, the pose of the object and the background vary, the efficiency of the standard classification methods drops significantly. Furthermore, if the positions and the number of the objects in the image are not given, the complexity of the problem increases. If the training set allows object segmentation, it is reasonable to employ this knowledge in the process of recognition training and reduce the influence of the background as much as possible. Several approaches are possible. We developed Hierarchical SVM method and compared it with Black-White method (BW) and Robust Principal Component Analysis (PCA). BW method is one of the pedagogical learning methods. In the learning phase the background pixels are replaced with extreme values. The hierarchical SVM and the robust PCA both rely on a subset of image pixels. Hierarchical SVM organizes the object pixels into hierarchical tree structure. In accordance with the hierarchical structure the hierarchical classifier is based solely on object pixels. In addition, PCA enables the reconstruction of the object in the recognition phase. The method starts with several hypothesis using different subsets of image points. With respect to reconstruction errors the points belonging to the background are first rejected and then the recognition proceeds. All three proposed methods are implemented and evaluated experimentally. The experimental results show that our Hierarchical SVM reduces false positive rate and maintains similar high recognition rate compared to the other two methods.

Actions (login required)