My Previous Research on Object Segmentation & Recognition
Dong Xu (Zhejiang Univerisity, China)
NICTA SML SEMINARDATE: 2005-11-16
TIME: 11:00:00 - 12:00:00
LOCATION: RSISE Seminar Room, ground floor, building 115, cnr. North and Daley Roads, ANU
CONTACT: JavaScript must be enabled to display this email address.
ABSTRACT:
To prevent moving cast shadows from being misunderstood as part of moving objects in change detection based video segmentation, we propose a novel approach to the cast shadow detection based on the edge and region information in multiple frames. First, an initial change detection mask containing moving objects and cast shadows is obtained. Then a Canny edge map is generated. After that, the shadow region is detected and removed through multi-frame integration, edge matching, and region growing. Finally, a post processing procedure is used to eliminate noise and tune the boundaries of the objects. Our approach can be used for video segmentation in indoor environment. The experimental results demonstrate its good performance.
Picture in Picture news video is one typical kind of video sequences, where some regions undergo the moving camera and the others undergo the static camera. Actually it is a very difficult for computer to distinguish whether the motion is caused by the moving objects themselves or the moving camera, thus extracting anchorperson from this kind of video becomes a challenging vision task. To our knowledge, it is the first time to address this kind of application. Change detection is implemented to achieve fast, automatic and robust anchorperson extraction with the aid of a simple anchorperson model, which is initialized once in the first frame. The experiment results demonstrate our performance.
In the traditional PCA/LDA and their kernel versions KPCA/KDA, the image matrix is concatenated into a one dimensional vector. Mostly the image object is represented in a very high dimensional feature space; whereas in some applications, such as face recognition, the available number of training samples is small, which usually results in the well known curse of the dimensionality and small sample problem. Furthermore, in the real world, the extracted feature of an object often has some specialized structures and such structures are in the form of 2nd or even higher-order tensors. For example, this is the case when a captured image is a 2nd order tensor, i.e. matrix, and when the sequence data such as a video for event analysis, are in the form of 3rd order tensor. So, by encoding an image object as a general tensor with 2nd or higher order, we propose a series work for dimension reduction, including unsupervised/supervised/multilinear/nonlinear versions: coupled subspace analysis (CSA-2), concurrent subspace analysis (CSA-3), bilinear discriminant analysis (BDA), discriminant analysis with tensor representation (DATER) and coupled kernel discriminant analysis (CKDA).
In the past decades, a large family of algorithmsa"€supervised or unsupervised; stemming from statistics or geometry theorya"€have been proposed to provide different solutions to the problem of dimensionality reduction. Beyond different motivations of these algorithms, we propose a general framework, graph embedding along with its linearization, kernelization and tensorization for dimensionality reduction, which reveals the underlying objective shared by most previous algorithms. It presents a unified perspective to understand these algorithms, that is, each algorithm can be considered as the direct graph embedding or its linear/kernel/tensor extension of some specific intrinsic graph characterizing certain desired statistic or geometry property of a data set, with the constraint from scale normalization or some penalty graph characterizing the statistic or geometry property that should be avoided. Furthermore, this framework can be used as a general platform to help develop new algorithms for dimensionality reduction. By taking graph embedding framework as a tool, we propose a new supervised algorithm, Marginal Fisher Analysis (MFA), for dimensionality reduction, in which the intrinsic graph characterizes the intra-class compactness and connects each data point with its neighboring points of the same class, while the penalty graph connects the marginal points and characterizes the inter-class separability. MFA effectively overcomes the limitations of traditional Linear Discriminant Analysis algorithm in terms of data distribution assumptions and available projection directions. The toy problem on artificial data and the real face recognition experiments both show the superiority of our proposed MFA in comparison to LDA, and also that of the kernel and tensor extensions of MFA over those of LDA.
BIO:
Dong XU is a final year Phd student of University of Science & Technology of China (USTC). He got the bachelor degree from USTC at 2001. He visited Microsoft Research Asia (MSRA) and the Chinese University of Hong Kong (CUHK). His research interests include Statistical Learning, Pattern Recognition, Computer Vision, Digital Image and Video Processing, and Human Computer Interaction.
