H. Quynh Dinh, Ph.D. q (at) hqdinh.com CV . Publications . Grants |
|
|
|
Research My research is in the areas of computer vision and graphics, with emphasis on shape reconstruction from point clouds, shape transformation, and geometric pattern matching. I also develop algorithms for shape and vector field pattern matching. The following summarizes my research projects. |
|
|
|
Vector Field Pattern Analysis | |
|
Analyzing vector fields impacts a diverse range of industries, including
computational fluid dynamics (CFD), ocean engineering, medical imaging, and
video and satellite imaging analysis among many others. We introduce algorithms
for detecting local and global patterns in vector fields by capturing the
statistics of the vector field in distributions that are then quantitatively
compared for pattern detection. We show that
global distributions
are capable of distinguishing between vector fields of
varying complexity and can be used to quantitatively compare similar fields.
To record local statistics, we adapt the spin-image representation for surface points to vector fields (shape contexts and other data structures can be similarly adapted). Local distributions are used to track points through vector fields with the ultimate goal of identifying the source of turbulent flow, a problem important to automotive and ocean engineers. In automotive engine design, a turbulent-free flow through the combustion chamber leads to optimal mixing of fuel and air and a more efficient combustion process. |
Publications:
Sponsor:
|
|
|
|
Analogical Search |
|
Search technology has become a significant market force due to the enormous amount of data that is publicly available on the internet. Data stores will only increase as more people contribute to media sharing sites (e.g., Flickr and Youtube), and this data will not be text data but rather, pictures, videos, and music. We are working with digital artists to develop a Transderivational/Transmedia search engine that suggests analogies across different media forms (e.g., images and videos) in a content-based manner. We are studying how visual elements such as shape, texture, and tone affect perceptions of similarity in images. The goal is to develop pattern matching algorithms that emulate the perceptual process (e.g., comparing the tone or structures within images) by which artists gather media samples for multimedia installations. | |
In working towards an analogical search engine, we focus on developing robust representations of geometric data and pattern matching within a single media form, including 3D shape matching and detecting patterns in vector fields derived from time-varying 2D and 3D datasets (see Vector Field Pattern Analysis above). For 3D shape matching, we develop a multi-resolution approach that generates a pyramid of geometric distributions to efficiently find corresponding points between 3D shapes. | |
Publications:
Sponsor:
|
|
|
|
Modeling Implicit Surfaces Composed of Radial Basis Functions |
|
In the digital entertainment industry, building geometric content for movies,
video games, and virtual worlds is labor-intensive, even with CAD tools. We
address this problem by smoothly interpolating discrete, noisy surface data
captured using consumer digital cameras. We develop an approach called volumetric regularization
to generate an implicit surface composed of
radial basis functions (RBF). Volumetric regularization uses energy-minimizing
3D RBFs to balance between data fitting and functional smoothness. Our method
generates a 3D implicit surface that approximates the data, closes off holes in
the data, and is locally detailed, yet globally smooth. To preserve sharp
features, we non-uniformly scale the RBFs, resulting in anisotropic RBFs.
Extensions to this work include evolving RBF parameters for segmenting images, interactive rendering and modeling of implicit surfaces on the GPU, and frequency-domain filtering and X-ray visualization of irregularly sampled volumes on the GPU. Filtering and zooming directly in frequency space leads to more accurate and efficient signal processing. By using anisotropic RBFs fitted to data through optimization techniques, we allow the inclusion of advanced data-sensitive constraints for feature preservation. |
|
Publications:
Sponsor:
|
|
|
|
Implicit Shape Transformation (Morphing) |
|
We reconstruct 4D implicit functions from time-varying surface data to obtain a 3D shape transformation (morph). Implicit shape transformation algorithms can create morphs between objects of arbitrary topology. However, explicit correspondences are not generated, making the transfer of surface properties (e.g., texture) impossible. Using heat diffusion, we build an explicit parameterization for implicit functions. In 4D, a parameterization provides a mapping between transforming 3D shapes and enables the transfer of material properties (e.g., color) from one surface to the next. | |
The observation of the evolution of a course of treatment can provide a powerful tool in understanding its efficacy. To visualize this, we produce animations allowing the visualization, as a function of time, of lesions in an organ. The animation produced is a morph describing how a source shape (pre-treatment) gradually deforms into a target shape (post-treatment). The morph is computed on the GPU, so both visualization of the volumes and morph generation are performed in real-time. |
|
Publications:
|
|
|
|
Thesis Advising |
|
|
|
|
|
College Courses Taught |
|
|
|
|