Lab Meeting January 19, 2010 (Jimmy): Semi-supervised Learning in Gigantic Image Collections

Title: Semi-supervised Learning in Gigantic Image Collections
Authors: Rob Fergus, Yair Weiss, and Antonio Torralba
In : NIPS2009

Abstract :
With the advent of the Internet it is now possible to collect hundreds of millions of images. These images come with varying degrees of label information. “Clean labels” can be manually obtained on a small fraction, “noisy labels” may be extracted automatically from surrounding text, while for most images there are no labels at all. Semi-supervised learning is a principled framework for combining these different label sources. However, it scales polynomially with the number of images, making it impractical for use on gigantic collections with hundreds of millions of images and thousands of classes. In this paper we show how to utilize recent results in machine learning to obtain highly efficient approximations for semi-supervised learning that are linear in the number of images. Specifically, we use the convergence of the eigenvectors of the normalized graph Laplacian to eigenfunctions of weighted Laplace-Beltrami operators. Our algorithm enables us to apply semi-supervised learning to a database of 80 million images gathered from the Internet.

[paper][slides]

Lab Meeting January 19, 2010(Chung-Han):A collection of outdoor robotic datasets with centimeter-accuracy ground truth

Title : A collection of outdoor robotic datasets with centimeter-accuracy ground truth
Author : Jose-Luis Blanco, Francisco-Angel Moreno, Javier-Gonzalez
In : Autonomous Robots (2009)

Abstract :
The lack of publicly accessible datasets with a reliable ground truth has prevented in the past a fair and coherent comparison of different methods proposed in the mobile robot Simultaneous Localization and Mapping (SLAM) literature. Providing such a ground truth becomes specially challenging in the case of visual SLAM, where the world model is 3-dimensional and the robot path is 6-dimensional. This work addresses both the practical and theoretical issues found while building a collection of six outdoor datasets. It is discussed how to estimate the 6-d vehicle path from readings of a set of three Real Time Kinematics (RTK) GPS receivers, as well as the associated uncertainty bounds that can be employed to evaluate the performance of SLAM methods. The vehicle was also equipped with several laser scanners, from which reference point clouds are built as a testbed for other algorithms such as segmentation or surface fitting. All the datasets, calibration information and associated software tools are available for download http://babel.isa.uma.es/mrpt/papers/dataset2009/.

[Full Text]

CMU PhD Thesis Defense: Mapping Large Urban Environments with GPS-Aided SLAM

CMU RI PhD Thesis Defense

Justin Carlson
Mapping Large Urban Environments with GPS-Aided SLAM
December 03, 2009, 9:00 a.m., GHC 6115

Abstract

Simultaneous Localization and Mapping (SLAM) has been an active area of research for several decades, and has become a foundation of indoor mobile robotics. However, although the scale and quality of results have improved markedly in that time period, no current technique can effectively handle city-sized urban areas.

The Global Positioning System (GPS) is an extraordinarily useful source of localization information. Unfortunately, the noise characteristics of the system are complex, arising from a large number of sources, some of which have large autocorrelation. Incorporation of GPS signals into SLAM algorithms requires using low-level system information and explicit models of the underlying system to make appropriate use of the information. The potential benefits of combining GPS and SLAM include increased robustness, increased scalability, and improved accuracy of localization.

This dissertation presents a theoretical background for GPS-SLAM fusion. The presented model balances ease of implementation with correct handling of the highly colored sources of noise in a GPS system.. This utility of the theory is explored and validated in the framework of a simulated Extended Kalman Filter driven by real-world noise.

The model is then extended to Smoothing and Mapping (SAM), which overcomes the linearization and algorithmic complexity limitations of the EKF formulation. This GPS-SAM model is used to generate a probabilistic landmark-based urban map covering an area an order of magnitude larger than previous work.

[LINK][PDF]

Thesis Committee
Charles Thorpe, Chair
Brett Browning
Martial Hebert
Frank Dellaert, Georgia Institute of Technology

Lab Meeting January 6th, 2010 (Any): Learning General Optical Flow Subspaces for Egomotion Estimation and Detection of Motion Anomalies

Author: Richard Roberts, Christian Potthast and Frank Dellaert

IEEE Computer Society Conference on Computer Vision and Pattern Recognition 2009 (CVPR'09)

Abstract--This paper deals with estimation of dense optical flow and ego-motion in a generalized imaging system by exploiting probabilistic linear subspace constraints on the flow. We deal with the extended motion of the imaging system through an environment that we assume to have some degree of statistical regularity. For example, in autonomous ground vehicles the structure of the environment around the vehicle is far from arbitrary, and the depth at each pixel is often approximately constant. The subspace constraints hold not only for perspective cameras, but in fact for a very general class of imaging systems, including catadioptric and multiple-view systems. Using minimal assumptions about the imaging system, we learn a probabilistic subspace constraint that captures the statistical regularity of the scene geometry relative to an imaging system. We propose an extension to probabilistic PCA (Tipping and Bishop, 1999) as a way to robustly learn this subspace from recorded imagery, and demonstrate its use in conjunction with a sparse optical flow algorithm. To deal with the sparseness of the input flow, we use a generative model to estimate the subspace using only the observed flow measurements. Additionally, to identify and cope with image regions that violate subspace constraints, such as moving objects, objects that violate the depth regularity, or gross flow estimation errors, we employ a per-pixel Gaussian mixture outlier process. We demonstrate results of finding the optical flow subspaces and employing them to estimate dense flow and to recover camera motion for a variety of imaging systems in several different environments.

Local copy

Lab Meeting January 6th, 2010 (Kuen-Han): Adaptive Relative Bundle Adjustment(RSS 2009)

Title: Adaptive Relative Bundle Adjustment(RSS 2009)
Authors: Gabe Sibley, Christopher Mei, Ian Reid, Paul Newman

Abstract

It is well known that bundle adjustment is the
optimal non-linear least-squares formulation of the simultaneous
localization and mapping problem, in that its maximum
likelihood form matches the definition of the Cramer Rao
Lower Bound. Unfortunately, computing the ML solution is
often prohibitively expensive – this is especially true during loop
closures, which often necessitate adjusting all parameters in a
loop. In this paper we note that it is precisely the choice of a single
privileged coordinate frame that makes bundle adjustment costly,
and that this expense can be avoided by adopting a completely
relative approach. We derive a new relative bundle adjustment,
which instead of optimizing in a single Euclidean space, works
in a metric-space defined by a connected Riemannian manifold.
Using an adaptive optimization strategy, we show experimentally
that it is possible to solve for the full ML solution incrementally
in constant time – even at loop closure. Our system also operates
online in real-time using stereo data, with fast appearance-based
loop closure detection. We show results for sequences of 23k
frames over 1.08km that indicate the accuracy of the approach.

paper link
technical paper

Lab Meeting December 30th, 2009 (Nicole) : Estimation of Sound Source Number and Directions under a Multi-source Environment (IROS 2009)

Title: Estimation of Sound Source Number and Directions under a Multi-source Environment (IROS 2009)

Authors: Jwu-Sheng Hu, Member IEEE, Chia-Hsing Yang, Student Member IEEE, and Cheng-Kang Wang

Abstract:

Sound source localization is an important featurein robot audition. This work proposes a sound source numberand directions estimation method by using the delayinformation of microphone array. An eigenstructure-basedgeneralized cross correlation method is proposed to estimatetime delay between microphones. Upon obtaining the time delayinformation, the sound source direction and velocity can beestimated by least square method. In multiple sound source case,the time delay combination among microphones is arrangedsuch that the estimated sound speed value falls within anacceptable range. By accumulating the estimation results ofsound source direction and using adaptive K-means++algorithm, the sound source number and directions can beestimated.

[link]

Measuring the Accuracy of Distributed Algorithms on Multi-Robot Systems

Measuring the Accuracy of Distributed Algorithms on Multi-Robot Systems

James McLurkin (UW CSE postdoc, MIT) 
October 9, 2008, 3:30 pm
EE-105

Abstract
Distributed algorithms running on multi-robot systems rely on ad-hoc networks to relay messages throughout the group. The propagation speed of these messages is large, but not infinite, and problems in algorithm execution can arise when the robot speed is a large fraction of the message propagation speed. This implies a robot speed limit, as any robot moving away from a message source faster than the message speed will never receive new information, and no algorithm can function properly on it. In this work, we focus on measuring the accuracy of multi-robot distributed algorithms. We define the Robot Speed Ratio (RSR) as the ratio of robot speed to message speed. We express it in a form that is platform-independent and captures the relationship between communications usage, robot mobility, and algorithm accuracy. We show that trade-offs between these key quantities can be balanced at design time. Finally, we present results from experiments with 50 robots that characterize the accuracy of preexisting distributed algorithms for network communication, navigation, boundary detection, and dynamic task assignment. In all cases, accuracy degrades as speed increases or communication bandwidth is reduced. In our experiments, a RSR of 0.005 allows good accuracy in all algorithms, a RSR of 0.02 allows reasonable accuracy in simple algorithms, and all algorithms tested are essentially useless at a RSR of 0.10 or higher.

[link]

Lab Meeting December 30th, 2009(Gary) : Pose Robust Face Tracking by Combining Active Appearance Models and Cylinder Head Models

Title: Pose Robust Face Tracking by Combining Active Appearance Models and Cylinder Head Models (IJCV 2008)

Author : Jaewon Sung , Takeo Kanade , Daijin Kim

Abstract:

The active appearance models (AAMs) provide
the detailed descriptive parameters that are useful for various
autonomous face analysis problems. However, they are
not suitable for robust face tracking across large pose variation
for the following reasons. First, they are suitable for
tracking the local movements of facial features within a limited
pose variation. Second, they use gradient-based optimization
techniques for model fitting and the fitting performance
is thus very sensitive to initial model parameters.
Third, when their fitting is failed, it is difficult to obtain
appropriate model parameters to re-initialize them. To alleviate
these problems, we propose to combine the active
appearance models and the cylinder head models (CHMs),
where the global head motion parameters obtained from the
CHMs are used as the cues of the AAM parameters for a
good fitting or re-initialization. The good AAM parameters
for robust face tracking are computed in the following manner.
First, we estimate the global motion parameters by the
CHM fitting algorithm. Second, we project the previously
fitted 2D shape points onto the 3D cylinder surface inversely
Third, we transform the inversely projected shape points by
the estimated global motion parameters. Fourth, we project
the transformed 3D points onto the input image and computed
the AAM parameters from them. Finally, we treat the
computed AAM parameters as the initial parameters for the
fitting. Experimental results showed that face tracking combining
AAMs and CHMs is more pose robust than that of
AAMs in terms of 170% higher tracking rate and the 115%
wider pose coverage.



link

NTU talk: Human Action Recognition Using Bag of Video Words

Title: Human Action Recognition Using Bag of Video Words
Speaker: Dr. Mubarak Shah, Agere Chair Professor of Computer Science, University of Central Florida
Time: 4:00pm, Dec 24 (Thu), 2009
Place: Room 210, CSIE building

Abstract:

The traditional approach for video analysis involves detection of objects, followed by tracking of objects from frame to frame and finally analysis of tracks for human action recognition. However, in some videos of complex scenes it is not possible to reliably detect and track objects. Therefore, recently in computer vision there has been lots of interest in the bag of video words approach, which bypasses the object detection and tracking steps. In bag of video words approach an action is described by a distribution of spatiotemporal cuboids (3D interest points).

In this talk, first I will describe a method to automatically discover the optimal number of video words clusters by utilizing the Maximization of Mutual Information (MMI). Unlike the k-means algorithm which is typically used to cluster spatiotemporal cuboids into video words based on their appearance similarity, MMI clustering further groups the video-words, such that the semantically similar video-words, e.g. words corresponding to the same part of the body during an action, are grouped in the same cluster.

The above method for human action recognition uses only one kind of features, spatiotemporal cuboids. However, single feature based representation for human action is not sufficient to capture the imaging variations (view-point, illumination etc.) and attributes of individuals (size, age, gender etc.).

Next I will present a method which uses two types of features: i) a quantized vocabulary of local spatio-temporal (ST) volumes (or cuboids), and ii) a quantized vocabulary of spin-images. To optimally combine these features, we treat different features and videos as nodes in a graph, where weighted edges between the nodes represent the strength of the relationship between entities. The graph is then embedded into a k-dimensional space subject to the criteria that similar nodes have Euclidian coordinates which are closer to each other. This is achieved by converting this constraint into a minimization problem whose solution is the eigenvectors of the graph Laplacian matrix. This procedure is known as Fiedler Embedding.

Short Biography:
Dr. Mubarak Shah, Agere Chair Professor of Computer Science, is the founding director of the Computer Visions Lab at UCF. He is a co-author of three books (Motion-Based Recognition (1997), Video Registration (2003), and Automated Multi-Camera Surveillance: Algorithms and Practice (2008)), all by Springer. He has published ten book chapters, seventy five journal and one hundred seventy conference papers on topics related to visual surveillance, tracking, human activity and action recognition, object detection and categorization, shape from shading, geo registration, photo realistic synthesis, visual crowd analysis, bio medical imaging, etc.

Dr. Shah is a fellow of IEEE, IAPR and SPIE. In 2006, he was awarded a Pegasus Professor award, the highest award at UCF, given to a faculty member who has made a significant impact on the university, has made an extraordinary contribution to the university community, and has demonstrated excellence in teaching, research and service. He is a Distinguished ACM Speaker. He was an IEEE Distinguished Visitor speaker for 1997-2000, and received IEEE Outstanding Engineering Educator Award in 1997. He received the Harris Corporation's Engineering Achievement Award in 1999, the TOKTEN awards from UNDP in 1995, 1997, and 2000; Teaching Incentive Program awards in 1995 and 2003, Research Incentive Award in 2003, Millionaires' Club awards in 2005 and 2006, University Distinguished Researcher award in 2007, SANA award in 2007, an honorable mention for the ICCV 2005 Where Am I? Challenge Problem, and was nominated for the best paper award in ACM Multimedia Conference in 2005. He is an editor of international book series on Video Computing; editor in chief of Machine Vision and Applications journal, and an associate editor of ACM Computing Surveys journal. He was an associate editor of the IEEE Transactions on PAMI, and a guest editor of the special issue of International Journal of Computer Vision on Video Computing. He is the program co-chair of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2008.

Lab Meeting Dez. 23rd 09 (Andi): Shape-based Recognition of 3D Point Clouds in Urban Environments

Authors: Aleksey Golovinskiy, Vladimir G. Kim, Thomas Funkhouser

International Conference on Computer Vision (ICCV), September 2009

This paper investigates the design of a system for recognizing objects in 3D point clouds of urban environments. The system is decomposed into four steps: locating, segmenting, characterizing, and classifying clusters of 3D points. Specifically, we first cluster nearby points to form a set of potential object locations (with hierarchical clustering). Then, we segment points near those locations into foreground and background sets (with a graph-cut algorithm). Next, we build a feature vector for each point cluster (based on both its shape and its context). Finally, we label the feature vectors using a classifier trained on a set of manually labeled objects. The paper presents several alternative methods for each step. We quantitatively evaluate the system and tradeoffs of different alternatives in a truthed part of a scan of Ottawa that contains approximately 100 million points and 1000 objects of interest. Then, we use this truth data as a training set to recognize objects amidst approximately 1 billion points of the remainder of the Ottawa scan.

full Paper

also:
Min-Cut Based Segmentation of Point Clouds
Aleksey Golovinskiy and Thomas Funkhouser
IEEE Workshop on Search in 3D and Video (S3DV) at ICCV, September 2009, Kyoto

Lab Meeting December 23rd, 2009 (Shao-Chen): Multi-robot SLAM with Unknown Initial Correspondence: The Robot Rendezvous Case

Title: Multi-robot SLAM with Unknown Initial Correspondence: The Robot Rendezvous Case (IROS 2006)

Authors: Xun S. Zhou and Stergios I. Roumeliotis

Abstract:

This paper presents a new approach to the multi-

robot  map-alignment  problem  that  enables  teams  of  robots  to

build joint maps without initial knowledge of their relative poses.

The  key  contribution  of  this  work  is  an  optimal  algorithm  for

merging  (not  necessarily  overlapping)  maps  that  are  created

by  different  robots  independently.  Relative  pose  measurements

between pairs of robots are processed to compute the coordinate

transformation  between  any  two  maps.  Noise  in  the  robot-

to-robot  observations,  propagated  through  the  map-alignment

process,  increases  the  error  in  the  position  estimates  of  the

transformed  landmarks,  and  reduces  the  overall  accuracy  of

the merged map. When there is overlap between the two maps,

landmarks that appear twice provide additional information, in

the form of constraints, which increases the alignment accuracy.

Landmark  duplicates  are  identified  through  a  fast  nearest-

neighbor  matching  algorithm.  In  order  to  reduce  the  compu-

tational  complexity  of  this  search  process,  a  kd-tree  is  used

to  represent  the  landmarks  in  the  original  map.  The  criterion

employed for matching any two landmarks is the Mahalanobis

distance.  As  a  means  of  validation,  we  present  experimental

results obtained from two robots mapping an area of 4,800 m

2.

[Link]

PhD Thesis Defense: Rhythmic Human-Robot Social Interaction

Marek P. Michalowski
Carnegie Mellon University
December 21, 2009, 10:00 a.m., NSH 3305

Abstract

Social scientists have identified and begun to describe rhythmic and synchronous properties of human social interaction. However, social interactions with robots are often stilted due to temporal mismatch between the behaviors, both verbal and nonverbal, of the interacting partners. This thesis brings the theory of interactional synchrony to bear on the design of social robots with a proposed architecture for rhythmic intelligence. We have developed technology that allows a robot to perceive social rhythms and to behave rhythmically. We have facilitated constrained social interactions, and designed experimental protocols, in which a robot variably synchronizes to human and/or environmental rhythms -- first in a dance-oriented task, and second in a cooperative video game. We have analyzed these interactions to understand the effects of Keepon's rhythmic attention on human performance. This thesis demonstrates that variations in a robot's rhythmic behavior have measurable effect on human rhythmic behavior and on performance in rhythmic tasks. Furthermore, human participants were able to assume and transition between the roles of leader or follower in these tasks.

Thesis Committee

Reid Simmons, Chair
Illah Nourbakhsh
Jodi Forlizzi
Hideki Kozima, Miyagi University, Japan

[link] [thesis draft]

Lab Meeting December 16th, 2009 (Casey): Monocular Vision SLAM for INdoor Aerial Vehicles

Title: Monocular Vision SLAM for INdoor Aerial Vehicles (IROS 2009)

Authors: Koray Celik, Soon-Jo Chung, Matthew Clausman, and Arun K. Somani

Abstract:

This paper presents a novel indoor navigation and ranging strategy by using a monocular camera. The proposed algorithms are integrated with simultaneous localization and mapping(SLAM) with a focus on indoor aerial vehicle applications. We experimentally validate the proposed algorithms by using a fully self-contained micro aerial vehicle (MAV) with on-board image processing and SLAM capabilities. The range measurement strategy is inspired by the key adaptive mechanisms for depth perception and pattern recognition found in humans and intelligent animals. The navigation strategy assumes an unknown, GPS-denied environment, which is representable via corner-like feature points and straight architectural lines. Experimental results show that the system is only limited by the capabilities of the camera and the availability of good corners.

[Link]

Lab Meeting December 16th, 2009 (Jeff): On measuring the accuracy of SLAM algorithms

Title: On measuring the accuracy of SLAM algorithms

Authors: Rainer Kümmerle, Bastian Steder, Christian Dornhege, Michael Ruhnke, Giorgio Grisetti, Cyrill Stachniss and Alexander Kleiner

Abstract:

In this paper, we address the problem of creating an objective benchmark for evaluating SLAM approaches. We propose a framework for analyzing the results of a SLAM approach based on a metric for measuring the error of the corrected trajectory. This metric uses only relative relations between poses and does not rely on a global reference frame. This overcomes serious shortcomings of approaches using a global reference frame to compute the error. Our method furthermore allows us to compare SLAM approaches that use different estimation techniques or different sensor modalities since all computations are made based on the corrected trajectory of the robot.
We provide sets of relative relations needed to compute our metric for an extensive set of datasets frequently used in the robotics community. The relations have been obtained by manually matching laser-range observations to avoid the errors caused by matching algorithms. Our benchmark framework allows the user to easily analyze and objectively compare different SLAM approaches.

Link:
Auton Robot 2009 27:387-407
http://www.springerlink.com/content/5u7458rl080216vr/fulltext.pdf

CMU Talk: Corridor View: Making Indoor Life Easier with Large Image Database

CMU VASC Seminar
Monday, Dec 7, 2009
1:30pm-2:30pm
NSH 1507

Corridor View: Making Indoor Life Easier with Large Image Database
Hongwen "Henry" Kang
Ph.D. Student, Robotics

Abstract:

Indoor environment poses substantial challenges for Computer Vision algorithms, due to the combined patterns that are either highly repetitive (e.g. doors), textureless (e.g. white walls), or temporally changing (e.g. posters, pedestrians). The fundamental challenge we want to tackle is the robust image matching. We proposed two approaches to address this problem, one is an iterative algorithm that combines global/local weighting strategies under bag-of-features model, the other data-mines distinctive feature vectors and uses high dimensional features directly for image matching, without quantization. Both of the approaches demonstrate significant improvements compared to straightforward image retrieval approaches, in highly confusing indoor environment. The proposed image matching techniques have broad applications. We selectively demonstrate two of them for this talk, specifically for vision impaired users living in the office environments. One application is data-driven zoomin; the other application is image composition for object pop-out.

CMU talk: Unsupervised Detection of Regions of Interest Using Iterative Link Analysis

CMU VASC Seminar
Monday, November 30, 2009

Unsupervised Detection of Regions of Interest Using Iterative Link Analysis
Gunhee Kim
Ph.D. Student, Computer Science Department

Abstract:

This work is a joint project with Antonio Torralba during my visit to MIT and will be presented as a poster at the upcoming NIPS 2009 Conference.

This talk will discuss a fast and scalable alternating optimization technique to detect regions of interest (ROIs) in cluttered Web images without labels. The proposed approach discovers highly probable regions of object instances by iteratively repeating the following two functions: (1) choose the exemplar set (i.e. a small number of highly ranked reference ROIs) across the dataset and (2) refine the ROIs of each image with respect to the exemplar set. These two subproblems are formulated as ranking in two different similarity networks of ROI hypotheses by link analysis. The experiments with the PASCAL 06 dataset show that our unsupervised localization performance is better than one of the state-of-the-art techniques and comparable to supervised methods. Also, we test the scalability of our approach with five objects in a Flickr dataset consisting of more than 200K images.

Bio: Gunhee Kim is a Ph.D. student in CMU's Computer Science Department advised by Takeo Kanade. He received his master's degree under the supervision of Martial Hebert in 2008 from the Robotics Institute at CMU. His research interests are computer vision, machine learning, data mining, and biomedical imaging.

Lab Meeting November 25, 2009 (Alan): Navigating, Recognizing and Describing Urban Spaces With Vision and Lasers (IJRR 2009)

Title: Navigating, Recognizing and Describing Urban Spaces With Vision and Lasers (IJRR 2009)

Authors: Paul Newman, Gabe Sibley, Mike Smith, Mark Cummins, Alastair Harrison, Chris Mei, Ingmar Posner, Robbie Shade, Derik Schroeter, Liz Murphy, Winston Churchill, Dave Cole, Ian Reid

Abstract:

In this paper we describe a body of work aimed at extending the reach of mobile navigation and mapping. We describe how running topological and metric mapping and pose estimation processes concurrently, using vision and laser ranging, has produced a full six degree-of-freedom outdoor navigation system. It is capable of pro-ducing intricate three-dimensional maps over many kilometers and in real time. We consider issues concerning the intrinsic quality of the built maps and describe our progress towards adding semantic labels to maps via scene de-construction and labeling. We show how our choices of representation, inference methods and use of both topological and metric techniques naturally allow us to fuse maps built from multiple sessions with no need for manual frame alignment or data association.

Link: http://ijr.sagepub.com/cgi/content/abstract/28/11-12/1406?etoc

CMU PhD Thesis Proposal: Learning Methods for Thought Recognition

CMU RI PhD Thesis Proposal

Mark Palatucci
Learning Methods for Thought Recognition
November 18, 2009, 3:00 p.m., NSH 3305

Abstract
This thesis proposal considers the problem of training machine learning classifiers in domains where data are very high dimensional and training examples are extremely limited or impossible to collect for all classes of interest. As a case study, we focus on the application of thought recognition, where the objective is to classify a person’s cognitive state from a recorded image of that person’s neural activity. Machine learning and pattern recognition methods have already made a large impact on this field, but most prior work has focused on classification studies with small numbers of classes and moderate amounts of training data. In this thesis, we focus on thought recognition in a limited data setting, where there are few, if any, training examples for the classes we wish to discriminate, and the number of possible classes can be in the thousands.

Despite these constraints, this thesis seeks to demonstrate that it is possible to classify noisy, high dimensional data with extremely few training examples by using spatial and temporal domain knowledge, intelligent feature selection, semantic side information, and large quantities of unlabeled data from related tasks.

In our preliminary work, we showed that it possible that build a binary classifier that can accurately classify between cognitive states with more than 80,000 features, and only two training examples per class. We also showed how classification can be improved using principled feature selection, and derived a significance test using order statistics that is appropriate for very high-dimensional problems with small numbers of training examples.

We have also explored the most extreme case of limited data, the zero-shot learning setting, where we do not have any training examples for classes we wish to discriminate. We showed that by using a knowledge base of semantic side information to create intermediate features, we can build a classifier that can classify words that people are thinking about, even without training data for those words while the classifier is forced to choose between nearly 1,000 different candidate words.

Finally, we showed how multi-task learning can be used to learn useful semantic features directly from data. We formulated the semantic feature learning problem as a Multi-task Lasso and presented an extremely fast and highly scalable algorithm for solving the resulting optimization.

We propose work to extend our zero-shot learning setting by optimizing semantic feature sets and by using an active learning framework to choose the most informative training examples. We also propose to use latent feature models such as components analysis and sparse coding in a self-taught learning framework to improve decoding by leveraging data from additional neural imaging experiments.

[PDF]

Thesis Committee
Tom Mitchell, Chair
Dean Pomerleau
J. Andrew Bagnell
Andrew Ng, Stanford University

CMU talk: Imitation Learning and Purposeful Prediction

Machine Learning Lunch (http://www.cs.cmu.edu/~learning/)
Speaker: Prof. Drew Bagnell
Date: Monday, November 23, 2009

Imitation Learning and Purposeful Prediction

Programming robots is hard. While demonstrating a desired behavior may be easy, designing a system that behaves this way is often difficult, time consuming, and ultimately expensive. Machine learning promises to enable "programming by demonstration" for developing high-performance robotic systems. Unfortunately, many approaches that utilize the classical tools of supervised learning fail to meet the needs of imitation learning. Perhaps foremost, classical statistics and supervised machine learning exist in a vacuum: predictions made by these algorithms are explicitly assumed to not affect the world in which they operate. I'll discuss the problems that result from ignoring the effect of actions influencing the world, and I'll highlight simple "reduction-based" approaches that, both in theory and in practice, mitigate these problems.

Additionally, robotic systems are often built atop sophisticated planning algorithms that efficiently reason far into the future; consequently, ignoring these planning algorithms in lieu of a supervised learning approach often leads to poor and myopic performance. While planners have demonstrated dramatic success in applications ranging from legged locomotion to outdoor unstructured navigation, such algorithms rely on fully specified cost functions that map sensor readings and environment models to a scalar cost. Such cost functions are usually manually designed and programmed. Recently, our group has developed a set of techniques that learn these functions from human demonstration by applying an /Inverse Optimal Control/ (IOC) approach to find a cost function for which planned behavior mimics an expert's demonstration. These approaches shed new light on the intimate connections between probabilistic inference and optimal control. I'll consider case studies in activity forecasting of drivers and pedestrians as well as the imitation learning of robotic locomotion and rough-terrain navigation. These case-studies highlight key challenges in applying the algorithms in practical settings.

Lab Meeting November 25, 2009 (KuoHuei): You’ll NeverWalk Alone: Modeling Social Behavior for Multi-target Tracking (ICCV 2009)

Title: You’ll NeverWalk Alone: Modeling Social Behavior for Multi-target Tracking

The Twelfth IEEE International Conference on Computer Vision (ICCV 2009)

Authors: S. Pellegrini, A. Ess, K. Schindler, and L. van Gool

Abstract:

Object tracking typically relies on a dynamic model to predict the object’s location from its past trajectory. In crowded scenarios a strong dynamic model is particularly important, because more accurate predictions allow for smaller search regions, which greatly simplifies data association.

Traditional dynamic models predict the location for each target solely based on its own history, without taking into account the remaining scene objects. Collisions are resolved only when they happen. Such an approach ignores important aspects of human behavior: people are driven by their future destination, take into account their environment, anticipate collisions, and adjust their trajectories at an early stage in order to avoid them. In this work, we introduce a model of dynamic social behavior, inspired by models developed for crowd simulation. The model is trained with videos recorded from birds-eye view at busy locations, and applied as a motion model for multi-people tracking from a vehicle-mounted camera. Experiments on real sequences show that accounting for social interactions and scene knowledge improves tracking performance, especially during occlusions.

[link]