Select a calendar:
Filter April Events by Event Type:
SUNMONTUEWEDTHUFRISAT
Events for April 08, 2014
-
PhD Defense - Zhenzhen Gao
Tue, Apr 08, 2014 @ 11:00 AM - 01:00 PM
Thomas Lord Department of Computer Science
University Calendar
PhD Candidate: Zhenzhen Gao
Title: City-scale Aerial LiDAR Point Cloud Visualization
Committee:
Ulrich Neumann (chair)
Aiichiro Nakano
C.-C. Jay Kuo (outside member)
Abstract:
Aerial LiDAR (Light Detection and Ranging) is cost-effective in acquiring terrain and urban information by mounting a downward-scanning laser on a low-flying aircraft. It produces huge volumes of unconnected 3D points. This thesis focuses on the interactive visualization of aerial LiDAR point clouds of cities, which is applicable to a number of areas including virtual tourism, security, land management and urban planning.
A framework needs to address several challenges in order to deliver useful visualizations of aerial LiDAR cities. Firstly, the data is 2.5D, in that the sensor is only able to capture dense details of the surfaces facing it, leaving few samples on vertical building walls. Secondly, the data often suffers from noise and undersampling. Finally, the large size of the data can easily exceed the memory capacity of a computer system.
This thesis first introduces a visually-complete rendering framework for aerial LiDAR cities. By inferring classification information, building walls and occluded ground areas under tree canopies are completed either through pre-processing point cloud augmentation or through online procedural geometry generation. A multi-resolution out-of-core strategy and GPU-accelerated rendering enable interactive visualization of virtually unlimited size data. With adding only a slight overhead to existing point-based approaches, the framework provides comparable quality to visualizations of off-line pre-computation of 3D polygonal models.
The thesis then presents a scalable out-of-core algorithm for mapping colors from aerial oblique imagery to city-scale aerial LiDAR points. Without intensive processing of points, colors are mapped via a modified visibility pass of GPU splatting, and a weighting scheme leveraging image resolution and surface orientation.
To alleviate visual artifacts caused by noise and under-sampling, the thesis shows an off-line point cloud refinement algorithm. By explicitly regularizing building boundary points, the algorithm can effectively remove noise, fill gaps, and preserve and enhance both normal and position discontinuous features for piecewise smoothing buildings with arbitrary shape and complexity.
Finally, the thesis introduces a new multi-resolution rendering framework that supports real-time refinement of aerial LiDAR cities. Without complex computation and without user interference, simply based on curvature analysis of points of uniform sized spatial partitions, hierarchical hybrid structures are constructed indicating whether to represent a partition as point or polygon. With the help of such structures, both rendering and refinement are dynamically adaptive to views and curvatures. Compared to visually-complete rendering, the new framework is able to deliver comparable visual quality with less than 8% increase in pre-processing time and 2-5 times higher rendering frame-rates. Experiments on several cities show that the refinement improves rendering quality for large magnification under real-time constraint.
Location: Charles Lee Powell Hall (PHE) - 333
Audiences: Everyone Is Invited
Contact: Lizsl De Leon
-
PhD Defense - Bin Liu
Tue, Apr 08, 2014 @ 02:00 PM - 04:00 PM
Thomas Lord Department of Computer Science
University Calendar
PhD Candidate: Bin Liu
Title: Improving Efficiency, Privacy and Robustness for Crowd-Sensing Applications
Committee:
Ramesh Govindan (chair)
Leana Golubchik
Sandeep Gupta (outside member)
Abstract:
Every year, a wide variety of modern smart devices, such as smartphones and tablets, are released by big brands, like Apple, Samsung and HTC. Compared to previous generations, these smart devices are more sophisticated in two ways: (a) they run advanced operating systems which allow developers to create a large collection of complicated apps, and (b) they have more diverse sensors which can be used to perform various context-aware tasks. These two attributes, together, have conceived a new class of applications, crowd-sensing. Crowd-sensing is a capability by which a task requestor can recruit smartphone users to provide sensor data to be used towards a specific goal or as part of a social or technical experiment. For the purpose of supporting crowd-sensing tasks, professional apps are developed to provide specialized platforms, and high quality sensors are used to generate semantically rich data.
My dissertation focuses on possible ways to improve efficiency, privacy and robustness for crowd-sensing applications. First, targeting the general form of crowd-sensing, we design efficient algorithms to answer the following question: how to optimize the selection of crowd-sensing participants to deliver credible information about a task? Based on a model about credibility of information, we develop solutions for the discrete version and the time-averaged version of this problem.
Second, we consider a special crowd-sensing case in which Internet-connected mobile users contribute sensor data as training samples, and collaborate on building a model for classification tasks such as activity or context recognition. Constructing the model can naturally be performed by a service running in the cloud, but users may be more inclined to contribute training samples if the privacy of these data could be ensured. For this, we develop algorithms and an associated system design to perform collaborative learning task in a way that preserves user data privacy without significant loss of accuracy.
Finally, the technique of dynamic analysis can be employed to test many aspects of crowd-sensing apps, such as performance, security, and correctness properties. As an initial attempt, we show how to use dynamic analysis to detect placement ad fraud in which app developers manipulate visual layouts of ads in ways that result in invisible ad impressions and accidental clicks from real users. We demonstrate that the detection can be performed using optimized automated navigation methods in a large set of 1,150 tablet apps and 50,000 phone apps.
Location: Henry Salvatori Computer Science Center (SAL) - 222
Audiences: Everyone Is Invited
Contact: Lizsl De Leon
-
Epstein Institute / ISE 651 Seminar Series
Tue, Apr 08, 2014 @ 03:30 PM - 04:50 PM
Daniel J. Epstein Department of Industrial and Systems Engineering
Conferences, Lectures, & Seminars
Speaker: Mengdi Wang, Department of Operations Research and Financial Engineering, Princeton University
Talk Title: "Stochastic Methods for Convex Optimization with 'Difficult' Constraints"
Abstract: Convex optimization problems involving large-scale data or expected values are challenging, especially when these difficulties are associated with the constraint set. We propose random algorithms for such problems, and focus on special structures that lend themselves to sampling, such as when the constraint is the intersection of many simpler sets, involves a system of inequalities, or involves expected values, and/or the objective function is an expected value, etc. We propose a class of new methods that combine elements of successive projection, stochastic gradient descent and proximal point algorithm. This class of methods also contain as special cases many known algorithms. We use a unified analytic framework to prove their almost sure convergence and the rate of convergence. Our framework allows the random algorithms to be applied with various sampling schemes (i.i.d, Markov, sequential, etc), which are suitable for applications involving distributed implementation, large data set, computing equilibriums, or statistical learning.
TUESDAY, APRIL 8, 2014
VON KLEINSMID CENTER (VKC) ROOM 100
3:30 - 4:50 PM
Biography: Mengdi Wang is currently an assistant professor at the Department of Operations Research and Financial Engineering, Princeton University. She received her PhD in Electrical Engineering and Computer Science from MIT in 2013. Her research interests include: large scale optimization, stochastic optimization algorithms, stochastic decision making, approximate dynamic programming, applications in statistical learning and finance, etc.
Host: Daniel J. Epstein Department of Industrial and Systems Engineering
More Information: Seminar-Wang_Mengdi.doc
Location: Von Kleinsmid Center For International & Public Affairs (VKC) - Room 100
Audiences: Everyone Is Invited
Contact: Georgia Lum
-
CS Colloquium: Abhishek Jain (MIT CSAIL): Computing on Private Data
Tue, Apr 08, 2014 @ 04:00 PM - 05:30 PM
Thomas Lord Department of Computer Science
Conferences, Lectures, & Seminars
Speaker: Abhishek Jain, MIT CSAIL
Talk Title: Computing on Private Data
Series: CS Colloquium
Abstract: THIS TALK WILL BE BROADCAST / STREAMING VIA THE FOLLOWING LINK. (Right click-open link in new tab or window.)
Today, end users generate large volumes of private data, some of which may be stored on the cloud in an encrypted form. The need to perform computation on this data to extract meaningful information has become ubiquitous.
The following fundamental questions arise in this setting: Can the cloud compute on the encrypted data of multiple users without knowing their secret keys? What functions can be computed in this manner? What if the users are mutually distrustful?
My research provides the first positive resolution of these questions. In this talk I will describe these new results and my other interests.
Biography: Abhishek Jain is currently a postdoctoral researcher in the Cryptography and Information Security Group at MIT CSAIL and Boston University. He received his PhD in Computer Science from UCLA in 2012 where he was the recipient of the Symantec Outstanding Graduate Student Research award. Abhishek's research interests are in cryptography and security, and related areas of theoretical computer science.
Host: Ming-Deh Huang
Location: Henry Salvatori Computer Science Center (SAL) - 101
Audiences: Everyone Is Invited
Contact: Assistant to CS chair
