Hsueh-Ti (Derek) Liu

About

I am a Mechanical Engineering MS student at Carnegie Mellon University, advised by Levent Burak Kara, Keenan Crane, and Alec Jacobson. My research focuses on 3D geometry processing using spectral and machine learning approaches. Previously I was an undergrad student at National Taiwan University majoring in Ocean Engineering, with research focused on ship design and fluid mechanics.
[Resume] [E-mail] [Github]


Research Interests

Spectral Shape Descriptors

Advisor: Keenan Crane, Alec Jacobson

The characterization and design of 3D shapes are central problems in computer graphics and geometric modelling. Nonetheless, a fundamental question - how different are the two shapes is still a challenging task. In the past few decades, many shape analysis tools have been developed. Spectral approaches, such as Laplace Operator, attracted huge attention because of the nice properties, such as rotation invariant, of spectral operators. However, Laplace Operator can only detect intrinsic shape properties. An important question is how to tell the extrinsic difference between shapes? This project aims to develop a new differential operator to solve this problem.

Data-driven Geometry Processing

Advisor: Levent Burak Kara

With the growth of 3D shape repositories, many data-driven geometry processing algorithms have been developed. Learning from shape collections can not only improve the performance of existing algorithms, such as shape correspondence, but also developing new algorithms based on the semantic meaning extracted from the data. For example, comfortability of a pair of shoes and sportiness of a given car shape. This research area aims to develop algorithms and applications based on the learned semantic features.

Data-driven Structural Optimization

Advisor: Levent Burak Kara

Structure optimization is a popular problem in 3D printing research. However, it is a time-consuming optimization process for a complex structure or for multiple loading conditions. In this research, we are exploring data-driven methods to speed-up the optimization process.


Projects

Deep Adversarial 3D Shape Net

10807: Topics in Deep Learning (Fall 2016)

3D shapes are a crucial but heavily underutilized resource in current computer vision research. Recently, deep learning models for 3D shapes have started to emerge because large amount of 3D shape data have become accessible. People are actively developing/finding proper 3D shape representations, such as uniform voxel representation and multi-view images. With the help of geometric approaches, this research topic tries to answer - is there a better way to represent 3D shapes for training deep learning models? [pdf]

Combining Active Learning and Accuracy Estimation using Unlabelled Data

10701: Introduction to Machine Learning (Spring 2016)

How unlabelled data can be used to estimate the true accuracy of learned classifier is an important problem across different fields. It is counter intuitive because we usually need labelled data to compute the true accuracy. In this project, we explored Error Estimation model which use graphical model approach to capture the dependency of making error across classifiers in order to estimate true accuracy using only unlabelled data. Then we brought this information to active learning model to verify the usefulness of estimated accuracy. [pdf]

Shape Descriptor Design Using Spin Transformation

15869: Discrete Differential Geometry (Spring 2016)

A well-known theory in differential geometry that mean curvature and metric should suffice to determine a surface uniquely. However, there has been no shape descriptor constructed using mean curvature and a metric. In this project, I applied Spin Transformation, mean curvature flow, and persistent barcode to construct a shape descriptor based on this theory. [pdf]

Recognition of Partially Obstructed 3D Objects from Point Cloud Data

24681: Computer-Aided Design (Spring 2016)

In the 3D printing industry, a metal printing process called the de-powdering process requires manually removing excess powder. In this project, we developed a framework to recognize 3D shapes from partially observed surface patches. With this framework in hand, we can then use robot arms to remove excess powder, which will significantly simplify the de-powdering process. [pdf]

Fundamental Fluid Mechanics Laboratory

Undergraduate Research (2013)

In this project, the main task is to explore the phenomenon when waves propagate over a submerged bar. We conducted wave propagation experiments and used a numerical model to simulate the experimental results.

Ship Model Basin Laboratory

Undergraduate Research (2013)

Optimizing hull positions of a trimaran through conducting ship model resistance tests. Based on the results, we optimized the hull positions and predicted the ship's performance in real size.