CSE 252C: Selected Topics in Vision and Learning

Project Reports

November 30, 2006

Boris Babenko: Task Specific Local Feature Matching

Abstract: Many problems in computer vision require the knowledge of potential point correspondences between two images. The usual approach for automatically determining correspondences begins by comparing small neighborhoods of high saliency in both images. Since speed is of the essence, most current approaches for local feature matching involve the computation of a feature vector that is invariant to various geometric and photometric transformations, followed by fast distance computations using standard vector norms. These algorithms include many parameters, and choosing an algorithm and setting its parameters for a given problem is more an art than a science. Furthermore, although invariance of the resulting feature space is in general desirable, there is necessarily a tradeoff between invariance and descriptiveness for any given task. In this paper we pose local feature matching as a classification problem, use powerful machine learning techniques to select a small number of simple features from a much larger pool, and train a classifier over these features. Our algorithm can be trained on specific domains or tasks, and performs better than the state of the art in such cases. Since our method is an application of the AdaBoost algorithm, we refer to it as Boosted Feature Matching (BOOM).

Nadav Ben-Haim: Real-time License Plate Recognition (LPR) and Make and Model Recognition (MMR)

Abstract: This is a technical project which explores the challenges of creating a mobile system to scan license plates and eventually recognize the make and model of a car as well. First, we explore the different kinds of cameras we can use and what kind works best. Depending on the kind of camera we have (IP, USB, Firewire, etc.) , there are different ways to interface with the camera via software. Illumination poses great difficulty in a mobile unit because of the variability involved. We study different LPR systems in use and try to understand what their approach is. For higher speeds, motion blur may become an issue depending on the kind of camera used. Setting parameters of the camera can help us resolve this issue to some extent. Also, since our processing occurs on a low resolution, we cannot use a generic approach to OCR for license plates. We introduce a simple, novel way to do OCR which exploits character width and baseline, but leave more sophisticated approaches for future work.

Adam Bickett: Catadioptric Stereo for Robot Localization

Abstract: Stereo rigs are indispensable in real world 3D localization and reconstruction, yet they are costly and the additional camera adds complexity. Catadioptric systems offer an inexpensive alternative, with the bonus of being taken from the same camera. I explore the design and calibration of such a rig, and begin to evaluate its utility for robot localization.

Marius Buibas: Determining Connectivity Structure of and Information Flow in Glial Neural Networks

Abstract: The use of the Optical Flow method is investigated for observation of intra-cellular Calcium signaling in neural glia. Image sequences are captured for cells incubated with a Ca2+ fluorescent dye, where increases in cytosolic calcium are indicated by an increase in image brightness. After the cells are mechanically stimulated, they communicate with other connected cells, propagating the change in calcium density and thus intensity. The optical flow technique is used to measure the speed of the signal and understand the cellular connectivity and receptor placement, by determining signal velocities at all points in the image sequence. We find that the signaling behaviour and functional connectivity is highly complex, showing a wide range of signal propagation speeds and patterns.

Carolina Galleguillos: Detecting and Recognizing Multiple Objects Using Context from Categories

Abstract: [put abstract here]

Tingfan Wu: Cafeteria Vision: Identification and Amount Measurement of Foods in a Plate

Abstract: We present a proto type of automatic dish recognition system, intended to ease the checkout process in self-serve cafeterias, where the price is based on what food and how much a customer takes. When a plate with several dishes in it goes though the checkout counter, a snapshot will be taken and then sent to a computer for analysis. A variety of features, including color and texture are extracted. Chi-square distance measure between color texton histograms is empirically proved to be best effective. Finally, support vector machines and nearest neighbor are used to classify the dishes against the pre-trained dish image bank. A preliminary system recognizing 24 dishes is constructed and several major difficulties are also identified.

Most recently updated on Oct. 29, 2006 by Serge Belongie.