CSE 252C: Selected Topics in Vision and Learning

Project Reports

December 2004

Manmohan Krishna Chandraker: Two-View Focal Length Estimation for All Camera Motions Using Priors

Abstract: Direct solutions to systems of polynomial equalities have been established in literature pertaining to two-view geometry to extract the focal lengths of a camera pair from the fundamental matrix, when other intrinsic calibration parameters are known. But these formulae break down near degeneracies characterized as critical motions, which occur quite commonly in practical scenarios. The contribution of this paper is twofold. First, a novel method of imposing priors on the calibration matrix is presented that allows us to formulate the problem of evaluating the focal length as a constrained minimization and arrive at a correct estimate even for degenerate configurations. Second, we explore solution techniques that achieve a global minimum for polynomial objective functions with feasible region constrained to be a semialgebraic set defined by polynomial equalities. Preliminary experimental results are presented on simulated noiseless data to demonstrate the validity of our theory, both in achieving a global minimum and doing so even for critical motions.

Louka Dlagnekov: License Plate Detection Using AdaBoost

Abstract: License Plate Recognition (LPR) is a fairly well explored problem with many successful solutions. Though most of these solutions are reasonably fast, there can be increased benefits to making them even faster, such as multiple recognition stages in video frames in a real-time video stream for improved accuracy. The goal of this project is to evaluate how well object detection methods used in text extraction and face detection apply to the problem of LPR. A strong classifier is trained by the AdaBoost algorithm and is used to classify parts of an image within a search window as either license plate or non-license plate.

Robin Hewitt: Learning to Generalize from a Small Number of Examples

Abstract: This report presents a method for detecting previously unseen instances of an object class in images by generalizing from a few representative examples of that object class. Detection proceeds from a set of initial hypotheses for the presence of the object class. The goodness of each hypothesis is evaluated by accumulation of evidence. Well-supported hypotheses are retained, while those that lack support are discarded. This method was implemented and tested at a proof-of-concept level by detecting a variety of previously unseen faces in cluttered environments after learning a generalized face model from just five images.

Sanjeev Kumar: Texture Synthesis and Image Inpainting for Video Compression

Abstract: In this report we investigate the application of texture synthesis and image inpainting techniques for video compression. Working in the non-parametric framework, we use 3D patches for matching and copying. This ensures temporal continuity to some extent which is not possible to obtain by working with individual frames. Since, in present application, patches might contain arbitrary shaped and multiple disconnected holes, fast fourier transform (FFT) and summed area table based sum of squared difference (SSD) calculation cannot be used. We propose a modification of above scheme which allows its use in present application. This results in significant gain of efficiency since search space is typically huge for video applications.

Hamed Masnadi-Shirazi: AdaBoost Face Detection

Abstract: Viola and Jones introduced a new and effective face detection algorithm based on simple features trained by the AdaBoost Algorithm, Integral Images and Cascaded Feature sets. This paper attempts to replicate their results. The Feret Face data set is used as the training set. The AdaBoost Algorithm, simple feature set and Integral Images are briefly explained and implemented in our Matlab based program. A series of ten best features were identified out of a set of close to fifty thousand. These best features were used to produce probability of error plots. Finally our face detection Algorithm is implemented on a series or random Images taken from the internet. More than just ten best features are needed to have a face detector comparable to the two hundred best features of Viola and Jones but the face detector still performs well and anyone can use our program included in the Appendix to implement this effective face detection algorithm and train as many best features as suited for their application.

Steve Scher: A Model of Perpendicular Texture for Determining Surface Geometry

Abstract: Three-dimensional objects distributed over a surface occlude each other in a manner dependent upon the angle (sigma) between the surface normal and the viewer. Leung & Malik note that occlusion allows the top of the objects to be seen, but not the bottom. For the case where the objects are differently colored as a function of height, the color content of the scene will thus depend upon sigma. A model is described that predicts the fraction of color expected to be present in each portion of the image, for a given value of sigma. This model can be compared to the actual color fractions in an image of unknown sigma, in order to furnish an error function, so that the sigma can be found using an optimization scheme. A set of synthetic images of cylinders on a plane are generated, and sigma is found to a mean accuracy of two degrees.

Rasit Onur Topaloglu: Forward Discrete Probability Propagation as an Alternative to Belief Propagation and Monte Carlo

Abstract: In a system formed of hierarchy of deterministic functions, estimation of the probability distribution of parameters through given root distributions is an algebraically intractable problem. This problem can be observed in a number of engineering applications, such as yield estimation of manufactured integrated chips or density estimation in image processing. This paper presents a novel approach that can handle arbitrary functions and distributions. The evaluation in this paper is done using linear functions of Gaussians. The comparison is given to a parametric belief propagation for Gaussians and a Monte Carlo based approach. The advantages of the proposed method over the compared methods have been presented.

Most recently updated on Nov. 28, 2004 by Serge Belongie.