Catherine Wah
Past and present projects.
Visipedia (UCSD, 2009-)
See here for more details.
Parking Space Vacancy Monitoring (UCSD, 2008)
Catherine Wah, Serge Belongie
Parking
The availability of parking spaces on campus at UCSD is a significant concern, and searching through crowded lots for available spots is frustrating, time consuming, and wastes gas. It would save time for the driver to be notified when spots are available, rather than have to search for them himself or herself.
In this project, we explore various vision-based methods for detecting vacancies in parking lots. In comparison to sensor-based parking space detection, which requires the installation and maintenance of networks of sensors, vision-based systems are more cost-effective and non-intrusive. We can monitor the vacancy status of parking spaces for the P502 parking lot on UCSD campus, using photos taken from the roof of CalIT2 with multiple pan-tilt-zoom (PTZ) cameras. In monitoring the parking lot, we must differentiate between A and B spaces.
Current methods of vacancy detection do not adequately deal with occlusion, particularly vehicle on vehicle, and this project seeks to address that issue. Other considerations that must be taken into account in the implementation of this vacancy monitoring system include variable lighting conditions, different orientations of the vehicle, and different poses of the vehicle.
Anthropometric Models to Estimate Hand Shape for Tracking (UIUC, 2008)
Catherine Wah, Dennis Lin, Thomas Huang
Hand Shape
Gestures are a natural form of interaction for humans, lending themselves to applications ranging from virtual reality systems to assisted communication. Accordingly, vision-based hand gesture recognition has shown promise as a computer interface medium, but the complexity and high dimensionality of hand shapes poses a barrier for effective modeling.
While previous work have explored methods for extracting a minimal feature set, we propose applying dimension reduction algorithms to compactly represent the natural variation found in human hands, as characterized primarily by internal hand features. We investigate various linear and nonlinear dimension reduction techniques, including generalized principal component analysis (GPCA), isometric feature mapping (Isomap), locally linear embedding (LLE), and maximum variance unfolding (MVU), each of which take a different geometric approach to arrive at the low dimensional representation. Starting with an initial set of 35 feature dimensions, we find an underlying embedding of 5-7 modes. From this low dimensional manifold, we then use radial basis function interpolations to reconstruct the higher dimensional parameters, which will be integrated into a tracking module of a gesture recognition system. Future work will involve observing improvements in tracking results.
The results of this research have particular importance in vision-based modeling and real-time tracking applications, where the number of complex objects being observed is usually exponential in the number of dimensions. Our observations can ultimately be applied to creating a system that can dynamically and automatically recognize a series of hand gestures, such as American Sign Language, to serve as a translator for the hearing impaired.
Attribute Inference Through Paired Comparisons (UCSD ECE 273, Spring 2010)
Catherine Wah
Similarity
We propose a novel approach for extracting attributes from tightly related object classes. Current work on attribute-based classification rely on pre-defined semantic attribute vocabularies, specified in advance by humans. Our method induces these attributes from a set of order relations of visual image similarities. This can be formulated as a convex optimization problem to find an optimal embedding of image similarity between subordinate categories.
From the generated embedding we can identify clusters or groupings of object classes that share similar visual features or attributes. For example, we can generate an embedding of Beak similarity for the Birds basic level category that encapsulates the variation in beak color, shape, size, etc. for Bird classes.
We find that the generated embeddings reflect the different modalities with which users judge similarity. These organically induced attributes capture visual similarity without being hindered by semantic labels and lend themselves to various vision and learning applications.
A Language for Interactive LED Visualization (UCSD CSE 231, Fall 2009)
Hayden Gomes, Emmett McQuinn, Catherine Wah
LEDs
The UCSD Computer Science and Engineering Department has built-in capability for LED displays, namely the interior programmable LED lights from Color Kinetics. These lights are located on each of the four floors of EBU3b, in the "Big Toe" portion of the building. Currently, these LEDs are rarely used and can only be manually controlled from a panel within the corresponding rooms. There are a select number of preprogrammed visualizations, but they are not easily modifable and do not support any direct or indirect interaction with the lights.
We propose a system for using these displays as a visualizer. For this project, we designed a language for controlling the LED displays. This visualization language is simple enough for members of the UCSD community to create their own visualizations using this language and then see their visualization displayed on the LED displays. We prototyped our system for the LEDs on a single floor, while enabling the visualization language to be extended to a full four-floor display. Our primary goal is to implement a system that allows the UCSD and CSE community to indirectly control the LED displays for an interactive experience.