Ishwar K. Sethi

Finding Trajectories of Feature Points in a Monocular Image Sequence

Identifying the same physical point in more than one image, the correspondence problem, is vital in motion analysis. Most research for establishing correspondence uses only two frames of a sequence to solve this problem. By using a sequence of frames, it is possible to exploit the fact that due to inertia the motion of an object cannot change instantaneously. By using smoothness of motion, it is possible to solve the correspondence problem for arbitrary motion of several nonrigid objects in a scene. We formulate the correspondence problem as an optimization problem and propose an iterative algorithm to find trajectories of points in a monocular image sequence. A modified form of this algorithm is useful in case of occlusion also. We demonstrate the efficacy of this approach considering synthetic, laboratory, and real scenes.

Entropy nets: from decision trees to neural networks

How the mapping of decision trees into a multilayer neural network structure can be exploited for the systematic design of a class of layered neural networks, called entropy nets (which have far fewer connections), is shown. Several important issues such as the automatic tree generation, incorporation of the incremental learning, and the generalization of knowledge acquired during the tree design phase are discussed. A two-step methodology for designing entropy networks is presented. The methodology specifies the number of neurons needed in each layer, along with the desired output, thereby leading to a faster progressive training procedure that allows each layer to be trained separately. Two examples are presented to show the success of neural network design through decision-tree mapping. >

Video shot detection and characterization for video databases

Abstract The organization of video information for video databases requires segmentation of a video into its constituent shots and their subsequent characterization in terms of content and camera work. In this paper, we look at these two steps using compressed video data directly. For shot detection, we suggest a scheme consisting of comparing intensity, row, and column histograms of successive I frames of MPEG video using the chi-square test. For characterization of segmented shots, we address the problem of classifying shot motion into different categories using a set of features derived from motion vectors of P and B frames of MPEG video. The central component of the proposed shot motion characterization scheme is a decision tree classifier built through a process of supervised learning. Experimental results using a variety of videos are presented to demonstrate the effectiveness of performing shot detection and characterization directly on compressed video.

Statistical approach to scene change detection

One of the challenging problems in video databases is the organization of video information. Segmenting a video into a number of clips and characterizing each clip has been suggested as one mechanism for organizing video information. This approach requires a suitable method to automatically locate cut points in a video. One way of finding such cut points is to determine the boundaries between consecutive camera shots. In this paper, we address this as a statistical hypothesis testing problem and present three tests to determine cut locations. All the three tests are such that they can be applied directly to the compressed video. This avoids an unnecessary decompression-compression cycle, since it is common to store and transmit digital video in compressed form. As our experimental results indicate, the statistical approach permits accurate detection of scene changes induced through straight as well as optical cuts.

Hierarchical Classifier Design Using Mutual Information

A nonparametric algorithm is presented for the hierarchical partitioning of the feature space. The algorithm is based on the concept of average mutual information, and is suitable for multifeature multicategory pattern recognition problems. The algorithm generates an efficient partitioning tree for specified probability of error by maximizing the amount of average mutual information gain at each partitioning step. A confidence bound expression is presented for the resulting classifier. Three examples, including one of handprinted numeral recognition, are presented to demonstrate the effectiveness of the algorithm.

Adaptive motion-vector resampling for compressed video downscaling

Digital video is becoming widely available in compressed form, such as a motion JPEG or MPEG coded bitstream. In applications such as video browsing or picture-in-picture, or in transcoding for a lower bit rate, there is a need to downscale the video prior to its transmission. In such instances, the conventional approach to generating a downscaled video bitstream at the video server would be to first decompress the video, perform the downscaling operation in the pixel domain, and then recompress it as, say, an MPEG, bitstream for efficient delivery. This process is computationally expensive due to the motion-estimation process needed during the recompression phase. We propose an alternative compressed domain-based approach that computes motion vectors for the downscaled (N/2xN/2) video sequence directly from the original motion vectors for the N/spl times/N video sequence. We further discover that the scheme produces better results by weighting the original motion vectors adaptively. The proposed approach can lead to significant computational savings compared to the conventional spatial (pixel) domain approach. The proposed approach is useful for video severs that provide quality of service in real time for heterogeneous clients.

Multimedia content processing through cross-modal association

Multimodal information processing has received considerable attention in recent years. The focus of existing research in this area has been predominantly on the use of fusion technology. In this paper, we suggest that cross-modal association can provide a new set of powerful solutions in this area. We investigate different cross-modal association methods using the linear correlation model. We also introduce a novel method for cross-modal association called Cross-modal Factor Analysis (CFA). Our earlier work on Latent Semantic Indexing (LSI) is extended for applications that use off-line supervised training. As a promising research direction and practical application of cross-modal association, cross-modal information retrieval where queries from one modality are used to search for content in another modality using low-level features is then discussed in detail. Different association methods are tested and compared using the proposed cross-modal retrieval system. All these methods achieve significant dimensionality reduction. Among them CFA gives the best retrieval performance. Finally, this paper addresses the use of cross-modal association to detect talking heads. The CFA method achieves 91.1% detection accuracy, while LSI and Canonical Correlation Analysis (CCA) achieve 66.1% and 73.9% accuracy, respectively. As shown by experiments, cross-modal association provides many useful benefits, such as robust noise resistance and effective feature selection. Compared to CCA and LSI, the proposed CFA shows several advantages in analysis performance and feature usage. Its capability in feature selection and noise resistance also makes CFA a promising tool for many multimedia analysis applications.

Feature point correspondence in the presence of occlusion

Occlusion and poor feature point detection are two of the main difficulties in the use of multiple frames for establishing correspondence of feature points. A formulation of the correspondence problem as an optimization problem is used to handle these difficulties. Modifications to an existing iterative optimization procedure for solving the formulation of the correspondence problem are discussed. Experimental results are presented to show the merits of the formulation. >

Direct feature extraction from compressed images

This paper examines the issue of direct extraction of low level features from compressed images. Specifically, we consider the detection of areas of interest and edges in images compressed using the discrete cosine transform (DCT). For interest areas, we show how a measure based on certain DCT coefficients of a block can provide an indication of underlying activity. For edges, we show using an ideal edge model how the relative values of different DCT coefficients of a block can be used to estimate the strength and orientation of an edge. Our experimental results indicate that coarse edge information from compressed images can be extracted up to 20 times faster than conventional edge detectors.

Machine recognition of constrained hand printed devanagari

Abstract A method is presented for the machine recognition of constrained, hand printed Devanagari characters. A set of very simple primitives is used, and all the Devanagari characters are looked upon as a concatenation of these primitives. Most of the decisions are taken on the basis of the presence/absence or positional relationship of these primitives; and the decision process is a multistage process, where each stage of decision making narrows down the choice regarding the class membership of the input token.