CMU Intelligence Seminar: An efficient way to learn deep generative models

An efficient way to learn deep generative models

Geoff Hinton, University of Toronto

I will describe an efficient, unsupervised learning procedure for deep generative models that contain millions of parameters and many layers of hidden features. The features are learned one layer at a time without any information about the final goal of the system. After the layer-by-layer learning, a subsequent fine-tuning process can be used to significantly improve the generative or discriminative performance of the multilayer network by making very slight changes to the features.

I will demonstrate this approach to learning deep networks on a variety of tasks including: Creating generative models of handwritten digits and human motion; finding non-linear, low-dimensional representations of very large datasets; and predicting the next word in a sentence. I will also show how to create hash functions that map similar objects to similar addresses, thus allowing hash functions to be used for finding similar objects in a time that is independent of the size of the database.

Speaker Bio

Geoffrey Hinton received his BA in experimental psychology from Cambridge in 1970 and his PhD in Artificial Intelligence from Edinburgh in 1978. He did postdoctoral work at Sussex University and the University of California San Diego and spent five years as a faculty member in the Computer Science department at Carnegie-Mellon University. He then became a fellow of the Canadian Institute for Advanced Research and moved to the Department of Computer Science at the University of Toronto. He spent three years from 1998 until 2001 setting up the Gatsby Computational Neuroscience Unit at University College London and then returned to the University of Toronto where he is a University Professor. He holds a Canada Research Chair in Machine Learning. He is the director of the program on "Neural Computation and Adaptive Perception" which is funded by the Canadian Institute for Advanced Research.

Geoffrey Hinton is a fellow of the Royal Society, the Royal Society of Canada, and the Association for the Advancement of Artificial Intelligence. He is an honorary foreign member of the American Academy of Arts and Sciences, and a former president of the Cognitive Science Society. He received an honorary doctorate from the University of Edinburgh in 2001. He was awarded the first David E. Rumelhart prize (2001), the IJCAI award for research excellence (2005), the IEEE Neural Network Pioneer award (1998) and the ITAC/NSERC award for contributions to information technology (1992).

[CMU RI Defense] Exploiting Space-Time Statistics of Videos for Face "Hallucination"

Title : Exploiting Space-Time Statistics of Videos for Face "Hallucination"

Author : Goksel Dedeoglu

Abstract :

Face "Hallucination" aims to recover high quality, high-resolution images of human faces from low-resolution, blurred, and degraded images or video. This thesis presents person-specific solutions to this problem through careful exploitation of space (image) and space-time (video) models. The results demonstrate accurate restoration of facial details, with resolution enhancements upto a scaling factor of 16.

The algorithms proposed in this thesis follow the analysis-by-synthesis paradigm; they explain the observed (low-resolution) data by fitting a (high-resolution) model. In this context, the first contribution is the discovery of a scaling-induced bias that plagues most model-to-image (or image-to-image) fitting algorithms. It was found that models and observations should be treated asymmetrically, both to formulate an unbiased objective function and to derive an accurate optimization algorithm. This asymmetry is most relevant to Face Hallucination: when applied to the popular Active Appearance Model, it leads to a novel face tracking and reconstruction algorithm that is significantly more accurate than state-of-the-art methods. The analysis also reveals the inherent trade-off between computational efficiency and estimation accuracy in low-resolution regimes.

The second contribution is a statistical generative model of face videos. By treating a video as a composition of space-time patches, this model efficiently encodes the temporal dynamics of complex visual phenomena such as eye-blinks and the occlusion or appearance of teeth. The same representation is also used to define a data-driven prior on a three-dimensional Markov Random Field in space and time. Experimental results demonstrate that temporal representation and reasoning about facial expressions improves robustness by regularizing the Face Hallucination problem.

The final contribution is an approximate compensation scheme against illumination effects. It is observed that distinct illumination subspaces of a face (each coming from a different pose and expression) still exhibit similar variation with respect to illumination. This motivates augmenting the video model with a low-dimensional illumination subspace, whose parameters are estimated jointly with high-resolution face details. Successful Face Hallucinations beyond the lighting conditions of the training videos are reported.

Full text of the thesis can be found here.
You can go here for more information.

Lab Meeting 26 April 2007 (AShin) :(ICRA 07)Outdoor Navigation of a Mobile Robot Using Differential GPS and Curb Detection

Title:Outdoor Navigation of a Mobile Robot Using Differential GPS and Curb Detection

Author:Seung-Hun Kim , Chi-Won Roh , Sung-Chul Kang and Min-Yong Park

Abstract— This paper demonstrates a reliable navigation ofa mobile robot in outdoor environment. We fuse differential GPS and odometry data using the framework of extended Kalman filter to localize a mobile robot. And also, we proposean algorithm to detect curbs through the laser range finder. Animportant feature of road environment is the existence of curbs.The mobile robot builds the map of the curbs of roads and themap is used for tracking and localization. The navigation systemfor the mobile robot consists of a mobile robot and a controlstation. The mobile robot sends the image data from a camera tothe control station. The control station receives and displays theimage data and the teleoperator commands the mobile robotbased on the image data. Since the image data does not containenough data for reliable navigation, a hybrid strategy forreliable mobile robot in outdoor environment is suggested.When the mobile robot is faced with unexpected obstacles or thesituation that, if it follows the command, it can happen to collide,it sends a warning message to the teleoperator and changes themode from teleoperated to autonomous to avoid the obstacles byitself. After avoiding the obstacles or the collision situation, themode of the mobile robot is returned to teleoperated mode. Wehave been able to confirm that the appropriate change ofnavigation mode can help the teleoperator perform reliablen avigation in outdoor environment through experiments in theroad.

CMU ML lunch: Learning without the loss function

Speaker: John Langford, Yahoo! Research, http://hunch.net/~jl/

Title: Learning without the loss function

Abstract: When learning a classifier, we use knowledge of the loss of different choices on training examples to guide the choice of a classifier. An often incorrect assumption is embedded in this paradigm: the assumption that we know the loss of different choices. This assumption is often incorrect, and the talk is about the feasibility of (and algorithms for) fixing this.

One example where the assumption is incorrect is the ad placement problem. Your job is to choose relevant ads for a user given various sorts of context information. You can test success by displaying an ad checking if the user is interested in it. However, this test does _not_reveal what would have happened if a different ad was displayed. Restated, the "time always goes forward" nature of reality does not allow us to answer "What would have happened if I had instead made a different choice in the past?"

Somewhat surprisingly, this is _not_ a fundamental obstacle to application of machine learning. I'll describe what we know about learning without the loss function, and some new better algorithms for this setting.

Lab Meeting 19 April 2007 (Nelson): Statistical Segment-RANSAC Localization and Moving Object Detection in Crowded Urban Area

• detailed algorithms
• working progress

Lab Meeting 19 May 2007 (ZhenYu) Stereovision with a Single Camera and Multiple Mirrors

Title : Stereovision with a Single Camera and Multiple Mirrors(ICRA2005)
Author : Mouaddib, E.M. Sagawa, R. Echigo, T. Yagi, Y.
Abstract:

You can create catadioptric omnidirectional stereovision using several mirrors with a single camera. These systems have interesting advantages, for instance in the case of mobile robot navigation and environment reconstruction. Our paper aims at estimating the” quality” of such stereovision system. What happens when the number of mirrors increases? Is it better to increase the base-line or to increase the number of mirrors? We propose some criteria and a methodology to compare different significant categories (seven): three already existing systems and four new designs that we propose. We also study and propose a global comparison between the best configurations.

[link]

SLAM product: a Web-based mapping system based on SRI International's SLAM technology

A message from SRI. -Bob

-----------------------------------
Hi.

We are pleased to announce the availability of a Web-based mapping system based on SRI International's SLAM technology called Karto™.

As part of the first step toward providing a complete suite of robotic navigation and exploration algorithms, we are making the Karto Logger Plug-In available as a free download at www.kartorobotics.com. The Karto Logger Plug-In allows robot developers to collect and log odometry and range data in a form that our mapping software can interpret to create maps of indoor environments. This can be done either in simulation or a on a real-world robotic platform. We currently support both the Player and Microsoft Robotic Studio platforms.

We welcome any and all feedback on our algorithms. Please feel free to contact us at kt-info@kartorobotics.com to tell us about your experiences with our mapping algorithm.

Sincerely,

The Karto™ Team
Software for Robot on the Move
http://www.kartorobotics.com

IEEE news: NEW PC SECURITY RECOGNIZES YOUR FACE

NEW PC SECURITY RECOGNIZES YOUR FACE

Forget passwords, Bioscrypt Inc. has recently introduced a USB-pluggable, 3-inch 3-D face recognition camera that can be used to authenticate users into computer systems. The new technology combines Bioscrypt's background as a provider of fingerprint-based biometric access controls with the advanced face imaging and recognition technology it acquired along with the A4Vision on March 14. Bioscrypt’s system will be called the VisionAccess 3D DeskCam and will work by casting a 40,000-point infrared mesh grid over the user’s face so that workers can log onto their computers, networks, and application with just a glance. For more information, visit: the link.

IEEE news: DRIVER-LESS SMART CARS TO HIT THE STREETS IN 15 YEARS

DRIVER-LESS SMART CARS TO HIT THE STREETS IN 15 YEARS

Once a pure figment of the Hollywood imagination, smart cars that operate autonomously are set to come to life in 15 years. Researchers at the University of Essex, Eastern England, are building a car using a standard remote control model, which will serve as a prototype for other researchers to develop their own smart cars. The cars will use a special type of computer software that will enable the car to recognize obstacles and make decisions. To read more, visit: the link.

[Lab meeting] 19 April 2007(Stanley)

I'll show the potential field method result.

[Machine Learning Lunchtime Chats]April 16, Pradeep Ravikumar, Sparsity recovery and structure learning

Title: Sparsity recovery and structure learning
Speaker: Pradeep Ravikumar

The sparsity pattern of a vector, the number and location of its non-zeros, is of importance in varied settings as structure learning in graphical models, subset selection in regression, signal denoising and compressed sensing. In particular, a key task in these settings is to estimate the sparsity pattern of an unknown ``parameter'' vector from a set of n noisy observations.

Suppose a sparse ``signal'' vector (edges in graphical models, covariates in regression) enters into linear combinations, and has observations which are functions of these linear combinations with added noise. The task is to identify the set of relevant signals from n such noisy observations. In graphical models for instance, the task in structure learning is to identify the set of edges from the noisy samples.

This is an intractable problem under general settings, but there has been a flurry of recent work on using convex relaxations and L1 penalties to recover the underlying sparse signal, and in particular, the sparsity pattern of the signal.

The tutorial will cover the basic problem abstraction, the applications in various settings and some general conditions under which the tractable methods ``work''. It will also focus in particular on the application setting of structure learning in graphical models.

ICRA '07 : Stereo-based Markerless Human Motion Capture for Humanoid Robot Systems

Title :
Stereo-based Markerless Human Motion Capture for Humanoid Robot Systems

Author :
Pedram Azad, Aleˇs Ude, Tamim Asfour, and R¨udiger Dillmann

Abstract :
In this paper, we present an image-based markerless human motion capture system, intended for humanoid robot systems. The restrictions set by this ambitious goal are numerous. The input of the system is a sequence of stereo image pairs only, captured by cameras positioned at approximately eye distance. No artificial markers can be used to simplify the estimation problem. Furthermore, the complexity of all algorithms incorporated must be suitable for real-time application, which is maybe the biggest problem when considering the high dimensionality of the search space. Finally, the system must not depend on a static camera setup and has to find the initial configuration automatically. We present a system, which tackles these problems by combining multiple cues within a particle filter framework, allowing the system to recover from wrong estimations in a natural way. We make extensive use of the benefit of having a calibrated stereo setup. To reduce search space implicitly, we use the 3D positions of the hands and the head, computed by a separate hand and head tracker using a linear motion model for each entity to be tracked. With stereo input image sequences at a resolution of 320×240 pixels, the processing rate of our system is 15 Hz on a 3 GHz CPU. Experimental results documenting the performance of our system are available in form of several videos.

ICRA 07: Rao-Blackwellized Particle Filtering for Mapping Dynamic Environments

Rao-Blackwellized Particle Filtering for Mapping Dynamic Environments

Isaac Miller and Mark Campbell


Abstract:

A general method for mapping dynamic environments
using a Rao-Blackwellized particle filter is presented.
The algorithm rigorously addresses both data association and
target tracking in a single unified estimator. The algorithm
relies on a Bayesian factorization to separate the posterior into
1) a data association problem solved via particle filter and
2) a tracking problem with known data associations solved
by Kalman filters developed specifically for the ground robot
environment. The algorithm is demonstrated in simulation and
validated in the real world with laser range data, showing
its practical applicability in simultaneously resolving data
association ambiguities and tracking moving objects.

Link

ICRA 07: Probabilistic Sonar Scan Matching for Robust Localization.

Probabilistic Sonar Scan Matching for Robust Localization.

Antoni Burguera, Yolanda Gonz´alez and Gabriel Oliver


Abstract:

This paper presents a probabilistic framework toperform scan matching localization using standard time-offlightultrasonic sensors. Probabilistic models of the sensors aswell as techniques to propagate the errors through the modelsare also presented and discussed. A method to estimate themost probable trajectory followed by the robot according tothe scan matching and odometry estimations is also presented.Thanks to that, accurate robot localization can be performedwithout the need of geometric constraints. The experimentsdemonstrate the robustness of our method even in the presenceof large amounts of noisy readings and odometric errors.

Link

ICRA 07: An accurate closed-form estimate of ICP’s covariance

An accurate closed-form estimate of ICP’s covariance

Andrea Censi

Abstract:
Existing methods for estimating the covariance of the ICP (Iterative Closest/Corresponding Point) algorithm are either inaccurate or are computationally too expensive to be used online. This paper proposes a new method, based on the analysis of the error function being minimized. It considers that the correspondences are not independent (the same measurement being used in more than one correspondence), and explicitly utilizes the covariance matrix of the measurements, which are not assumed to be independent either. The validity of the approach is verified through extensive simulations: it is more accurate than previous methods and its computational load is negligible. The ill-posedness of the surface matching problem is explicitly tackled for under-constrained situations by performing an observability analysis; in the analyzed cases the method still provides a good estimate of the error projected on the observable manifold.

LINK

ICRA 07: Improved Data Association for ICP-based Scan Matching in Noisy and Dynamic Environments

Improved Data Association for ICP-based Scan Matching in Noisy and Dynamic Environments

Diego Rodriguez-Losada and Javier Minguez

Abstract:

This paper presents a technique to improve the data association in the Iterative Closest Point based scan matching. The method is based on a distance-filter constructed on the basis of an analysis of the set of solutions produced by the associations in the sensor configuration space. This leads to a robust strategy to filter all the associations that do not explain the principal motion of the scan (due to noise in the sensor, large odometry errors, spurious, occlusions or dynamic features for example). The experimental results suggest that the improvement of the data association leads to more robust and faster methods in the presence of wrong correspondences.

LINK

Lab Meeting 12 May 2007 (fish60)

Just talk a little about the progress of the hand.

Lab Meeting 12 May 2007 (Vincent): Face recognition Using Probabilistic Eigenface

In this talk, I will present a face recognition approach that integrates eigenface with probabilistic aprroach. I'll also demonstrate some result using this approach.

This work is based on the following paper.

Title : Beyond eigenfaces: probabilistic matching for face recognition

Author : Moghaddam, B. Wahid, W. Pentland, A.

Abstract :
We propose a technique for direct visual matching for face recognition and database search, using a probabilistic measure of similarity which is based on a Bayesian analysis of image differences. Specifically we model lure mutually exclusive classes of variation between facial images: intra-personal (variations in appearance of the same individual, due to different expressions or lighting) and extra-personal (variations in appearance due to a difference in identity). The likelihoods for each respective class are learned from training data using eigenspace density estimation and used to compute similarity based on the a posteriori probability of membership in the intra-personal class, and ultimately used to rank matches in the database. The performance advantage of this probabilistic technique over nearest-neighbor eigenface matching is demonstrated using results front ARPA's 1996 “FERET” face recognition competition, in which this algorithm was found to be the top performer.

Here is the full text.

MERTZ: An active vision head robot for exploring social learning

MERTZ is an active vision head robot, designed for exploring scalable learning in a social context. Inspired by how human infants learn by observing and imitating other people, we plan to have MERTZ be placed in a public venue for long periods of time, continuously interacting with people and incrementally learning about various correlations. For example, the robot may learn to correlate objects and people with frequently uttered phoneme sequences, differentiate among people and their interaction habits, learn to dislike some people who tend to annoy the robot, etc. MERTZ has recently gone through a series of experiment where it interacted with many people at different public spaces in the Stata Center.

related links: MERTZ projects , Lijin Aryananda ,
Research Abstracts ,
PhD thesis: A Few Days of A Robot's Life in the Human's World: Toward Incremental Individual Recognition

VASC Seminar Series

Piotr Dollar
UC San Diego
Monday, April 9, 3:30pm, NSH 1507

Supervised Learning of Edges and Object Boundaries

Edge detection is one of the most studied problems in computer vision, yet it remains a very challenging task. It is difficult since often the decision for an edge cannot be made purely based on low level cues such as gradient, instead we need to engage all levels of information, low,middle, and high, in order to decide where to put edges. In this paper we propose a novel supervised learning algorithm for edge and object boundary detection which we refer to as Boosted Edge Learning or BEL for short. A decision of an edge point is made independently at each location in the image; a very large aperture is used providing significant context for each decision. In the learning stage, the algorithm selects and combines a large number of features across different scales in order to learn a discriminative model using an extended version of the Probabilistic Boosting Tree classification algorithm. The learning based framework is highly adaptive and there are no parameters to tune. We show applications for edge detection in a number of specific image domains as well as on natural images. We test on various datasets including the Berkeley datasetand the results obtained are very good.

Short Bio:
Piotr is currently pursuing his PhD in Computer Science at UCSD, under the guidance of Prof. Serge Belongie (expected Sept 2007). He obtained his B.A and M.S degrees from Harvard University. His interests lie in machine learning and pattern recognition, and their application to computer vision. His research is supported by the NSF IGERT fellowship.

[MIT technical report] A Few Days of A Robot’s Life in the Human's World: Toward Incremental Individual Recognition

Title : A Few Days of A Robot’s Life in the Human’s World: Toward Incremental Individual Recognition

Author : Lijin Aryananda

Abstract :

This thesis presents an integrated framework and implementation for Mertz, an expressive robotic creature for exploring the task of face recognition through natural interaction in an incremental and unsupervised fashion. The goal of this thesis is to advance toward a framework which would allow robots to incrementally “get to know” a set of familiar individuals in a natural and extendable way. This thesis is motivated by the increasingly popular goal of integrating robots in the home. In order to be effective in human-centric tasks, the robots must be able to not only recognize each family member, but also to learn about the roles of various people in the household. In this thesis, we focus on two particular limitations of the current technology. Firstly, most of face recognition research concentrate on the supervised classification problem. Currently, one of the biggest problems in face recognition is how to generalize the system to be able to recognize new test data that vary from the training data. Thus, until this problem is solved completely, the existing supervised approaches may require multiple manual introduction and labelling sessions to include training data with enough variations. Secondly, there is typically a large gap between research prototypes and commercial products, largely due to lack of robustness and scalability to different environmental settings. In this thesis, we propose an unsupervised approach which would allow for a more adaptive system which can incrementally update the training set with more recent data or new individuals over time. Moreover, it gives the robots a more natural social recognition mechanism to learn not only to recognize each person’s appearance, but also to remember some relevant contextual information that the robot observed during previous interaction sessions. Therefore, this thesis focuses on integrating an unsupervised and incremental face recognition system within a physical robot which interfaces directly with humans through natural social interaction. The robot autonomously detects, tracks, and segments face images during these interactions and automatically generates a training set for its face recognition system. Moreover, in order to motivate robust solutions and address scalability issues, we chose to put the robot, Mertz, in unstructured public environments to interact with naive passersby, 2 instead of with only the researchers within the laboratory environment. While an unsupervised and incremental face recognition system is a crucial element toward our target goal, it is only a part of the story. A face recognition system typically receives either pre-recorded face images or a streaming video from a static camera. As illustrated an ACLU review of a commercial face recognition installation, a security application which interfaces with the latter is already very challenging. In this case, our target goal is a robot that can recognize people in a home setting. The interface between robots and humans is even more dynamic. Both the robots and the humans move around. We present the robot implementation and its unsupervised incremental face recognition framework. We describe an algorithm for clustering local features extracted from a large set of automatically generated face data. We demonstrate the robot’s capabilities and limitations in a series of experiments at a public lobby. In a final experiment, the robot interacted with a few hundred individuals in an eight day period and generated a training set of over a hundred thousand face images. We evaluate the clustering algorithm performance across a range of parameters on this automatically generated training data and also the Honda-UCSD video face database. Lastly, we present some recognition results using the self-labelled clusters.

Here is the full text

Stanford Close To A.I. Robots For Home

Stanford is getting close to creating a robot that can help you around the house. Japan is also so close that government officials released a set of ethical guidelines to protect people from robots.

"Stair" is Stanford's artificial intelligence robot project that can be given a simple task, such as, empty the dishwasher. Then analysis it's surroundings and then figures out the best way to complete it, on its own.

Link (with video for full content)

News: Aircraft Swarm Around Single Airborne Controller

NewScientist.com news service , Paul Marks
17:40 02 April 2007

Fighter pilots will one day be able to control entire squadrons of uncrewed combat aircraft as well as their own plane, following successful flight demonstrations of a multi-aircraft remote control system in UK airspace.
In addition to cutting the number of pilots risked in military operations, the remote control system could one day also be used to auto-land hijacked planes. Or they might allow lone pilots to orchestrate complex search and rescue operations.
UK defence firm Qinetiq demonstrated the system on 30 March. The pilot of a modified Tornado fighter plane assumed remote control of a BAC 1-11 airliner carrying members of the press, including New Scientist, and flying at an altitude of 4500 metres (15000 feet). The Tornado pilot was also in control of three simulated Unmanned Combat Air Vehicles (UCAVs).
Currently, UCAVs and their unarmed cousins, UAVs, are controlled remotely by pilots on the ground, who may be thousands of miles away. Autonomy could allow an airborne fighter pilot closer to the action take control, with each UCAV sending images and other data back to the control jet.
Autonomy has more control than a normal autopilot system and is more sophisticated, coordinating the movement of several different airplanes simultaneously.
It does this by assigning a software agent to look after each UCAV or UAV and automatically drawing up flight patterns around the likely targets. The uncrewed craft follow these patterns until the fighter pilot, who examines images of possible targets, decides they should investigate, attack, or go home to refuel.

<Original Link, Video>

CMU ML talk: Mining Large Time-evolving Data Using Matrix and Tensor Tools

Christos Faloutsos, CMU
Tamara G. Kolda, Sandia National Labs
Jimeng Sun, CMU

How can we find patterns in sensor streams (eg., a sequence of temperatures, water-pollutant measurements, or machine room measurements)? How can we mine Internet traffic graph over time? Further, how can we make the process incremental? We review the state of the art in four related fields: (a) numerical analysis and linear algebra (b) multi-linear/tensor analysis (c) graph mining and (d) stream mining. We will present both theoretical results and algorithms as well as case studies on several real applications. Our emphasis is on the intuition behind each method, and on guidelines for the practitioner.

The link.

IEEE RASeNews

Volume 5 No. 4
April 4 2007

Welcome to RASeNews, the IEEE RAS email newsletter. RASeNews is limited to short announcements of RAS Conferences, RAS society 'action items', breaking news, and time-sensitive issues. Send comments, queries and news items to r.g.snyder@ieee.org. If you received this newsletter from a colleague and are interested in joining the IEEE Robotics and Automation Society, go to the link.

CONTENTS

1. ICRA Travel grants for female graduate students
2. RAS Membership Forum at ICRA 2007
3. Petitions for RAS AdCom
4. IROS Deadline Extended to April 9
5. Other Imminent Deadlines
***********

Last Minute ICRA Travel Grants

On April 3 the ICRA committee was informed that Microsoft Research will provide $5000 for travel grants for female grad students to attend ICRA and the Women's Lunch/BOF. Microsoft Research wants the awards to go to students that would not otherwise be able to attend. For more information, contact Nancy Amato, amato @ cs.tamu.edu.

-----------
RAS Open Forum at ICRA 2007

At their October 2006 meeting, the AdCom voted to have the Open Forum be comprised of reports from the major committees and boards that would normally be given to the AdCom at their formal meeting after ICRA. This will give members a chance to see in detail what the AdCom does and provide inputs to the AdCom regarding decisions and initiatives which the Adcom will consider on Sunday, April 15. (All RAS members who wish to attend the AdCom meeting are welcome.).
This year the Open Forum will feature the President's report, reports on our Publications, the Conference Editorial Board and the Electronic Products and Services Board. The Forum will be held beginning at 6:20 p.m. (18:20, immediately after the last technical session) on Wednesday, April 11 in Angelicum A, Aula Minor. Please plan to attend.


-----------
Petition for IEEE-RAS AdCom Election

The slate of candidates for the RAS Administrative Committee is normally chosen by the AdCom nominations committee, currently chaired by Art Sanderson (sandea@rpi.edu), Toshio Fukuda (fukuda @ mein.nagoya-u.ac.jp) and Kazuo Tanie (ktanie@cc.tmit.ac.jp). However, a candidate who submits a petition with the signatures of at least 2 percent of RAS voting members (Graduate Student, Members, Affliate, Senior and Fellows) he/she is automatically included on the ballot. Petitions must be received by August 15 to be guaranteed a place on the ballot. To obtain a petition form and more information, contact the RAS Society Administrator (r.g.snyder@ieee.org). Also, according to the RAS Standing Rules, the Nominations Committee will seriously consider any candidate who obtains at least 25 signatures on his/her petition. The nominations committee welcomes nominations, including self nominators.

-------------------------
IROS 2007 Extended
Due to a number of requests, the Program committee of the Intelligent Robots and Systems Conference has extended the submissions deadline to 9 April.
http://www.iros2007.org
IROS 2007, San Diego. 29 Oct- 2 Nov

---------------------
Other Imminent Deadlines
CONFERENCES
IEEE CASE 2007 22-25 Sept. DEADLINE 30 April 2007.
http://www.fulton.asu.edu/~case2007

MMAR 27-30 August, 2007. Szczecin, Poland. DEADLINE 16 April, 2007.
http://www.mmar.ps.pl/

PUBLICATIONS
IEEE-TASE Special Issue on Automation and Engineering for Ambient Intelligence. Submissions Deadline 1 May 2007.
IEEE Transactions on Automation Science and Engineering.
http://www.ieee-ras.org/tase/files/CFPAmbientIntelligence.pdf

IEEE RAM Special Issue on Design, Control, and Applications of Real-World Multi-Robot Systems.
Submissions Deadline 1 June 2007.
IEEE Robotics and Automation Magazine
http://www.ieee-ras.org/ram/special.php#mrs

*******

IEEE news: HOLLYWOOD IMAGINATIONS JUMPS OUT INTO REAL LIFE

Resembling the system Tom Cruise used in Steven Spielberg’s film Minority Report, Gesture Studios has developed a system they call GoodPoint using a technology called motion capture, which film studios and video game makers have used for years to make computer-animated characters appear more realistic. The interface featured in Minority Report is just what Gesture is selling to companies using cameras to track hand movements and translate them into computer instructions to create presentations at the 14,000 trade shows and conferences in the U.S. each year. The technology is now bursting out of Hollywood and changing the way consumers interact with home electronics. The technology has also been adopted by advertisers to create interactive displays. To read more, visit: the link.

IEEE news: ‘ALICE’ AIMS TO HELP GIRLS BUILD PROGRAMMING SKILLS

Alice, a Java-based interactive program that allows users to produce 3-D computer animation images without the need for advanced level programming skills has been created by a lab director at Carnegie Mellon University (USA) in an effort to draw middle-school age girls to the computer science field. Through the use of a computer mouse, young users can create stories, outline the actions of characters or objects, and perform graphics programs using simple commands, such as a drop and drag interface. CMU officials claim Alice has been successful in teaching programming skills to female students from middle school to college-age. The link.

IEEE news: GPS NAVIGATION TO HELP BLIND PEOPLE

Il Village, a northern Italy technology company, has developed the Easy Walk service, which will give blind people greater independence and mobility. Easy Walk uses a mobile phone, a small Bluetooth GPS receiver, text to speech software, and a call center that will operate seven days a week. The Easy Walk is scheduled to be launched in the fall. Read more, the link.

Unsupervised Learning of Boosted Tree Classifier using Graph Cuts for Hand Pose Recognition

Title: Unsupervised Learning of Boosted Tree Classifier using Graph Cuts for Hand Pose Recognition
Author: Toufiq Parag, Ahmed Elgammal
(BMVC 2006)

Abstract:
This study proposes an unsupervised learning approach for the task of hand pose recognition.
Considering the large variation in hand poses, classification using a decision tree seems highly
suitable for this purpose. Various research works have used boosted decision trees and have
shown encouraging results for pose recognition. This work also employs a boosted classifier
tree learned in an unsupervised manner for hand pose recognition. We use a recursive two way
spectral clustering method, namely the Normalized Cut method (NCut), to generate the decision
tree. A binary boosting classifier is then learned at each node of the tree generated by the clustering algorithm. Since the output of the clustering algorithm may contain outliers in practice, the variant of boosting algorithm applied at each node is the Soft Margin version of AdaBoost, which was developed to maximize the classifier margin in a noisy environment. We propose a novel approach to learn the weak classifiers of the boosting process using the partitioning vector given by the NCut algorithm. The algorithm applies a linear regression of feature responses with the partitioning vector and utilizes the sample weights used in boosting to learn the weak hypotheses. Initial result shows satisfactory performances in recognizing complex hand poses with large variations in background and illumination. This framework of tree classifier can also be applied to general multi-class object recognition.

paper download:[link]

CMU vasc talk: The role of Manifold learning in Human Motion Analysis

Ahmed Elgammal
Rutgers, State University of New Jersey
Monday, April 2, 3:30PM, NSH 1507

The role of Manifold learning in Human Motion Analysis

Human body is an articulated object with high degrees of freedom. Despite the high dimensionality of the configuration space, many human motion activities lie intrinsically on low dimensional manifolds. Although the intrinsic body configuration manifolds might be very low in dimensionality, the resulting appearance manifolds are challenging to model given various aspects that affects the appearance such as the shape and appearance of the person performing the motion, or variation in the view point, or illumination. Our objective is to learn representations for the shape and the appearance of moving (dynamic) objects that support tasks such as synthesis, pose recovery, reconstruction, and tracking. We studied various approaches for representing global deformation manifolds that preserve their geometric structure. Given such representations, we can learn generative models for dynamic shape and appearance. We also address the fundamental question of separating style and content on nonlinear manifolds representing dynamic objects.We learn factorized generative models that explicitly decompose the intrinsic body configuration (content) as a function of time from the appearance/shape (style factors) of the person performing the action as time-invariant parameters. We show results on pose recovery, body tracking, gait recognition, as well as facial expression tracking and recognition.

News: Shoulder-worn camera acts as a third eye

NewScientist.com news service , Tom Simonite

A shoulder-mounted camera system that automatically tracks head movements and can recognise hand gestures has been developed by UK researchers. Eventually, they hope the system could identify a wearer's activity and offer assistance, for example by accessing a telephone directory when they reach for the phone.

The collar-mounted camera is worn on one shoulder. It is controlled wirelessly from a laptop computer, which uses the camera's output to keep track of objects, map its position and recognise different hand gestures made by the user. Walterio Mayol Cuevas, now a researcher at Bristol University, UK, created the camera while working at the University of Oxford.

Three separate motors make the camera highly directional, and even allow it to tilt, while inertia sensors are used to keep it pointed correctly while the wearer moves around

video: wearable camera, VSLAM
full article
One of the UK researcher: W.W. Mayol

News : Arms are dead giveaway for risky drivers

• 31 March 2007
• NewScientist.com news service
• Paul Marks

MONITORING a driver's gaze - as some luxury cars are now designed to do - may not be the best way to ensure they are paying attention to the road ahead. A more effective way to spot distracted drivers may be to monitor their activity by tracking head, arm and hand movements.

In the US, around 25 per cent of road accidents are caused by inattentive drivers, according to the National Highway Traffic Safety Administration. To encourage drivers to keep their eyes on the road, cars such as the Lexus 450 are now fitted with twin cameras mounted behind the steering wheel to monitor the direction of their gaze. If they look away from the road for too long an alarm sounds.

But in a paper to be published in the journal Transportation Research C (tinyurl.com/2983ug), engineers at the University of Minnesota, Minneapolis, say eye gaze is a poor measure because people can be distracted without taking their eyes off the road. "People don't always look down at the dashboard to adjust the aircon or the music system. And you can talk on a cellphone while still looking ahead," says researcher Harini Veeraraghavan. Gaze detectors can also be fooled by sunglasses, says his supervisor Nikos Papanikolopoulos.

Instead, the researchers have developed a system that looks at a driver's overall actions and then classifies them as safe or unsafe, sounding an alarm if necessary. An intelligent camera continually monitors the driver and tracks what they are doing with their head, arms and hands by picking out their skin tone from the background. This data is fed into software that has learned from thousands of examples which combinations of head, arm and hand movements could be risky - for example, holding a hand to the face for long periods to eat or talk on a cellphone.

In tests, the system has shown promise, although it needs refining. Clothing with skin-coloured patches can fool the software, for instance. Sebastian Enders, a researcher in driver distraction at the German in-car entertainment firm Blaupunkt, says the approach is interesting. "Their challenge will be proving that the system can reliably detect risky behaviour time after time," he says.

From issue 2597 of New Scientist magazine, 31 March 2007, page 28

Origin

CMU ML lunch: A unifying view of component analysis (from a computer vision perspective)

Speaker: Fernando De la Torre, CMU
http://www.cs.cmu.edu/~ftorre
Title: A unifying view of component analysis (from a computer vision perspective)
Date: April 02


Abstract:
Component Analysis (CA) methods (e.g. Kernel Principal Component Analysis, Independent Component Analysis, Tensor factorization) have been used as a feature extraction step for modeling, classification and clustering in numerous visual, graphics and signal processing tasks over the last four decades. CA techniques are especially appealing because many can be formulated as eigen-problems, offering great potential for efficient learning of linear and non-linear representations of the data without local minima.

In the first part of the talk, I will review standard CA techniques (Principal Component Analysis, Canonical Correlation Analysis, Linear Discrimiant Analysis, Non-negative Matrix Factorization, Independent Component Analysis) and three standard extensions (Kernel methods, latent variable models and tensors). In the second part of the talk, I will describe a unified framework for energy-based learning in CA methods. I will also propose several extensions of CA methods to learn linear and non-linear representations of data to improve performance, over the current use of CA features, in state-of-the-art algorithms for classification (e.g. support vector machines), clustering (e.g. spectral graph methods) and modeling/visual tracking (e.g. active appearance models) problems.

(FYI) Position at NASA Ames: Google Earth Content Developer

GOOGLE EARTH CONTENT DEVELOPER

The Intelligent Robotics Group at the NASA Ames Research Center has an immediate opening for a full-time software engineer. We are looking for an developer with experience in computer vision to help publish NASA's images and 3D models of the Earth, Moon, and Mars to millions of people using Google Earth. The ideal candidate will have knowledge and experience with image processing, image stitching, and/or 3D vision.

Responsibilities:

* Work with research scientists and content specialists to create and publish a new NASA layer on Google Earth.
* Develop efficient, high-performance image processing pipelines.
* Implement rectification, alignment, and stitching methods for very large image datasets, including oblique angle imagery.
* Implement projection and mosaicing methods for very large 3D terrain models (Digital Elevation Models).

Requirements:
* B.S. (or higher) in Computer Science, with at least 2 years experience in team-based software development projects.
* At least 2 years experience with image processing/computer vision toolkits such as OpenCV.
* Significant development experience in C++ and UNIX environment (GNU tools, svn, doxygen).
* Knowledge of stereo vision and/or GIS is preferred.

NOTE: U.S. citizenship or current permanent resident status is REQUIRED.

The NASA Ames Intelligent Robotics Group (http://ti.arc.nasa.gov/irg) builds systems to help humans explore and understand extreme environments, remote locations and uncharted worlds. IRG conducts applied research in a wide range of areas including: real-time 3D graphics, computer vision, gigapixel panoramic imaging, human-robot interaction, robot software architectures and planetary rovers. NASA Ames is located in the heart of Silicon Valley and is a world leader in information technology for space.

If you are interested in applying for this position, please send the following via email:
- a letter describing your background and motivation
- a detailed resume (text or PDF format)
- contact details for two (or more) references
to Dr. Terry Fong .

Lab Meeting 29 March 2007 (Jeff): RFID only SLAM Simulator

Briefly talking about the RFID SLAM simulation
1.Motion Model
2.RFID Sensor Model
3.Show the odometry only and SLAM simulation

News: IF YOU'RE HAPPY, THE ROBOT KNOWS IT

I hang my head and sink into my chair dejectedly. As I slouch, the computer monitor in front of me tilts forward and drops low to almost touch the desk, mimicking my gloomy posture. When I perk up and straighten my back, the computer spots the change and the monitor cheerfully swings forward and upward.

Meet RoCo, the world's first expressive computer. Inhabiting a back room in MIT Media Lab, the robotic computer has a monitor for a head and a simple LCD screen for a face. It expresses itself using its double-jointed neck, which is equipped with actuators that shift the monitor up and down, tilt it forward and back and swivel it from side to side, rather like Pixar's animated lamp. An attached camera can detect when its user moves, allowing RoCo to adjust its posture accordingly.

RoCo's creators hope that by responding to a user's changes in posture, people might be more likely to build up a "rapport" with the computer that will make sitting at a desk all day a little more enjoyable. The MIT researchers also believe that by tuning into users' moods, the robot might help them get their work done more effectively.

See the full article and video
From issue 2596 of New Scientist magazine, 22 March 2007, page 30-31

CMU RI seminar:Exposing Digital Forgeries from Inconsistencies in Lighting

Hany Farid
Professor, Computer Science Department
Dartmouth College

With the advent of high-resolution digital cameras, powerful personal computers and sophisticated photo-editing software, the manipulation of digital images is becoming more common. To this end, we have been developing a suite of tools to detect tampering in digital images. I will discuss two related techniques for exposing forgeries from inconsistencies in lighting. In each case we show how to estimate the direction to a light source from only a single image: inconsistencies across the image are then used as evidence of tampering.
[Joint work with Kimo Johnson]

Lab Meeting 22 March 2007 (Nelson): Statistical Segment-RANSAC Localization and Moving Object Detection in Crowded Urban Area

• a new scheme for improving RANSAC ICP
• detailed algorithms
• working progress
  

IEEE news: FIRM DEBUTS WIRELESS VIDEO-CONFERENCING ROBOT

FIRM DEBUTS WIRELESS VIDEO-CONFERENCING ROBOT

The Giraffe Video-Conferencing Robot, a 5-foot-8 mobile video-conferencing system, can roam an office like a surrogate supervisor, according to the firm HeadThere. The device’s 14-inch screen and 2-megapixel camera can look up and down and turn in place to face speakers in a meeting. HeadThere says its robot can be lowered so that its "head" is at sitting height for face-to-face conversation. The system will be available in 2008 at a price between $1,800 and $3,000. To read more, visit: the link.

FIRM DEBUTS BRAINWAVE SOFTWARE TO CONTROL APPLICATIONS

FIRM DEBUTS BRAINWAVE SOFTWARE TO CONTROL APPLICATIONS
Software that reads thoughts and emotions through sensors and uses the data to control a computer has been developed by the San Francisco-based firm Emotiv, which says its technology reads both conscious and non-conscious brainwaves. Developer kits include a lightweight headset with multiple sensors and a wireless transmitter, Emotiv says, and its three software products can identify facial expressions, measure users’ emotional states, and allow them to use brainwaves to move or manipulate objects by thinking about an action, such as push, pull, lift, or rotate. Emotiv says it expects to have consumer products for gaming consoles and PC games based on its technology available in 2008. Read more: the link.

IEEE news: SENSORS MADE TO DETECT HAZARDOUS UNDERWATER ENVIRONMENTS

SENSORS MADE TO DETECT HAZARDOUS UNDERWATER ENVIRONMENTS

A research team at the University of Illinois at Urbana-Champaign has created a sensor that mimics the lateral line of some underwater species that helps them survive in underwater environments. This artificial lateral line, made for use with submarines and underwater vehicles, will detect and track moving underwater targets and avoid collisions with moving or stationary objects. The artificial lateral line consists of an integrated linear array of micro fabricated flow sensors, with the sizes of individual sensors and spacings between them matching those of their biological counterpart. For more information, visit: Link

CMU FRC talk: Particle RRT for Path Planning on Very Rough Terrain

Particle RRT for Path Planning on Very Rough Terrain

Nik Melchior. Ph.D. Candidate. Robotics Institute

Wednesday, March 28, 2007

Autonomous navigation algorithms operate on a model of the world built using information perceived by the robot. Despite quantifiable uncertainties in perception and modelling techniques, many navigation algorithms use coarsely quantified evaluations of areas of the world. Often, areas are classified as safe/unsafe or low cost/high cost, when much richer information is available to the planner.
This work introduces a path planning technique which explicitly models the uncertainty in the terrain and its properties. The method is an extension to the Rapidly-exploring Random Tree (RRT) algorithm, a well-known approach to path planning with kinodynamic constraints in high-dimensional state spaces.
Our extension, called Particle RRT (pRRT), uses multiple simulations to propagate the estimated uncertainty in perception to the planned path itself. This allows us to plan paths which are significantly more robust, and which can be followed with greater accuracy, even using open-loop control.

Lab Meeting 22 March 2007 (Bright): Continuum Crowds

SIGGRAPH 2006

Author: Adrien Treuille, Seth Cooper and Zoran Popovi´c

Abstract:

We present a real-time crowd model based on continuum dynamics.In our model, a dynamic potential field simultaneously integratesglobal navigation with moving obstacles such as other people, efficientlysolving for the motion of large crowds without the need forexplicit collision avoidance. Simulations created with our systemrun at interactive rates, demonstrate smooth flow under a varietyof conditions, and naturally exhibit emergent phenomena that havebeen observed in real crowds.

Link

Lab Meeting 22 March 2007 (YuChun): Effects of Anticipatory Action on Human-Robot Teamwork

HRI 2007

Author:
Guy Hoffman and Cynthia Breazeal

Abstract:
A crucial skill for fluent action meshing in human team activity is a learned and calculated selection of anticipatory actions. We believe that the same holds for robotic teammates, if they are to perform in a similarly fluent manner with their human counterparts.
In this work, we propose an adaptive action selection mechanism for a robotic teammate, making anticipatory decisions based on the confidence of their validity and their relative risk. We predict an improvement in task efficiency and fluency compared to a purely reactive process.
We then present results from a study involving untrained human subjects working with a simulated version of a robot using our system. We show a significant improvement in best-case task efficiency when compared to a group of users working with a reactive agent, as well as a significant difference in the perceived commitment of the robot to the team and its contribution to the team’s fluency and success. By way of explanation, we propose a number of fluency metrics that differ significantly between the two study groups.

Link

Lab Meeting 22 March 2007 (Jim): A visual bag of words method for interactive qualitative localization and mapping

A visual bag of words method for interactive qualitative localization and mapping
David Filliat, ICRA 2007

PDF, homepage

Abstract:
Localization for low cost humanoid or animal-like personal robots has to rely on cheap sensors and has to be robust to user manipulations of the robot. We present a visual localization and map-learning system that relies on vision only and that is able to incrementally learn to recognize the different rooms of an apartment from any robot position. This system is inspired by visual categorization algorithms called bag of words methods that we modified to make fully incremental and to allow a user-interactive training. Our system is able to reliably recognize the room in which the robot is after a short training time and is stable for long term use. Empirical validation on a real robot and on an image database acquired in real environments are presented.


Cited paper -- Visual Categorization with Bags of Keypoints.

Lab Meeting 22 Mar 2007 (Atwood): Fast, Integrated Person Tracking and Activity Recognition with Plan-View Templates from a Single Stereo Camera

Author:
Michael Harville, Hewlett-Packard Laboratories Center for Signal and Image Processing
Dalong Li, Georgia Inst. of Technology

Abstract:
Plan-view projection of real-time depth imagery can improve
the statistics of its intrinsic 3D data, and allows for
cleaner separation of occluding and closely-interacting people.
We build a probabilistic, real-time multi-person tracking
system upon a plan-view image substrate that well preserves
both shape and size information of foreground objects.
The tracking’s robustness derives in part from its “plan-view
template” person models, which capture detailed properties
of people’s body configurations. We demonstrate that these
same person models - obtained with a single compact stereo
camera unit - may also be used for fast recognition of body
pose and activity. Principal components analysis is used to
extract plan-view “eigenposes”, onto which person models,
extracted during tracking, are projected to produce a compact
representation of human body configuration. We then
formulate pose recognition as a classification problem, and
use support vector machines (SVMs) to quickly distinguish
between, for example, different directions people are facing,
and different body poses such as standing, sitting, bending
over, crouching, and reaching. The SVM outputs are transformed
to probabilities and integrated across time in a probabilistic
framework for real-time activity recognition.

Link

[Robot Perception and Learning] Lab Meeting Day

Since Bob won't go to ARTC this Thursday, the lab meeting will be rescheduled back to 11 AM ~2PM, this Thursday.

CFP: Microsoft Robotics Studio Soccer Challenge at RoboCup 2007 Atlanta

http://www.robocup-us.org/microsoft

Microsoft Robotics Studio Soccer Challenge at RoboCup 2007 Atlanta

Call for Participants

Deadline: March 19, 2007

Through this call, RoboCup 2007 Atlanta requests a Letter of Intent from those teams interested in participating in a series of competitive soccer matches using the Microsoft Robotics Studio physics-based 3D simulation environment at RoboCup 2007 Atlanta.

Microsoft is creating a robot soccer simulation to demonstrate how Microsoft Robotics Studio can be applied to challenging environments like RoboCup soccer. While this contest will demonstrate simulation of a competitive event, Microsoft is also working with robot manufacturers developing hardware for RoboCup to demonstrate how easily player competition software can be easily transferred to actual robots.

Information about the event: Participating teams will use a special preview of Microsoft new soccer competition. The initial version of the competition will feature wheeled robots available for free download from the Microsoft Robotics Studio website: http://www.microsoft.com/robotics. Teams can use this to begin initial development of their player software.

By May, one or more advanced simulated models (i.e. legged robots) are expected to be available for the soccer simulation. One of these simulated robots will be selected for the final competition and released to the community at that time. Teams can then migrate their code to the new robot in preparation for the challenge in July.

Recognition: Placing teams will be awarded trophies and will be recognized on the Microsoft and RoboCup 2007 Atlanta websites.

How to respond to this call: Responses should include the following information:

Team Name
Team Leader
University (or company) Name
Country
Short description of your previous participation in RoboCup

Responses should be sent by email to tucker@cc.gatech.edu

CMU vision talk: A Logical Theory for Detecting Humans in Surveillance Video (with an excursion into collaborative filtering)

A Logical Theory for Detecting Humans in Surveillance Video (with an excursion into collaborative filtering)

LARRY S. DAVIS, University of Maryland
_______________________
> March 19, 2007
> 3:30 PM
> Hamerschlag Hall 1112
> Refreshments 3:15 PM
________________________

The capacity to robustly detect humans in video is a critical component of automated visual surveillance systems. This talk describes a bilattice based logical reasoning approach that exploits contextual information, and knowledge about interactions between humans, and augments it with the output of low level body part detectors for human detection. Detections from low level parts-based detectors are treated as logical facts and used to reason explicitly about the presence or absence of humans in the scene. Positive and negative information from different sources, as well as uncertainties from detections and logical rules, are integrated within the bilattice framework. This approach also generates proofs or justifications for each hypothesis it proposes. These justifications (or lack thereof) are further employed by the system to explain and validate, or reject potential hypotheses. This allows the system to explicitly reason about complex interactions between humans and handle occlusions. These proofs are also available to the end user as an explanation of why the system thinks a particular hypothesis is actually a human. We employ a boosted cascade of gradient histograms based detector to detect individual body parts. We have applied this framework to analyze the presence of humans in static images from different datasets.

I will also talk about the application of this framework to the problem of collaborative filtering. In these applications, there exist, potentially, a large number of cues that can contribute to making a final recommendation for a given user. The proposed bilattice based logical reasoning approach exploits these multiple, noisy, and potentially contradictory sources of information to predict movie preferences. We report results on the publicly available MovieLens dataset and compare our approach against a number of state-of-the-art ranking algorithms for collaborative filtering.

CMU Intelligence Seminar: Statistical Relational Learning: Entity Resolution and Link Prediction

Statistical Relational Learning: Entity Resolution and Link Prediction

Lise Getoor
University of Maryland - College Park

A key challenge for machine learning is mining richly structured datasets describing objects, their properties, and links among the objects. We'd like to be able to learn models which can capture both the underlying uncertainty and the logical relationships in the domain. Links among the objects may demonstrate certain patterns, which can be helpful for many practical inference tasks and are usually hard to capture with traditional statistical models. Recently there has been a surge of interest in this area, fueled in part by interest in mining social networks, web collections, security and law enforcement data and biological data.

Statistical Relational Learning (SRL) is a newly emerging research area which attempts to represent, reason and learn in domains with complex relational and rich probabilistic structure. In this talk, I'll begin with a short SRL overview. Then, I'll describe some of my group's recent work, focusing on our work on entity resolution and link prediction in relational domains.

Joint work with students: Indrajit Bhattacharya, Mustafa Bilgic, Rezarta Islamaj, Louis Licamele, Galileo Namata, Vivek Sehgal, Prithviraj Sen and Elena Zheleva.

Speaker Bio
Prof. Lise Getoor is an assistant professor in the Computer Science Department at the University of Maryland, College Park. She received her PhD from Stanford University in 2001. Her current work includes research on link mining, statistical relational learning and representing uncertainty in structured and semi-structured data. Her work in these areas has been supported by NSF, NGA, KDD, ARL and DARPA. In June 2006, she co-organized the fourth in a series of successful workshops on statistical relational learning, http://www.cs.umd.edu/projects/srl2006/. She has published numerous articles in machine learning, data mining, database and AI forums. She was one of 11 finalists choosen nationally for the 2005 Microsoft New Faculty Award. She is a member of AAAI Executive council, is on the editorial board of the Machine Learning Journal, is a JAIR associate editor and has served on numerous program committees including AAAI, ICML, IJCAI, KDD, SIGMOD, UAI, VLDB, and WWW.

Lab Meeting 15 Mar 2007 (Any): RFID-Based SLAM

Urban Search And Rescue (USAR) is a time critical task. Extraordinary circumstances after a real disaster make it very hard to apply common techniques. Firemen at 911 reported that they had major difficulties to orientate themselves after leaving collapsed buildings. The arbitrary structure of the environment and limited visibility conditions due to smoke, dust, and fire, prevent an easy distinction of different places. Therefore, it is proposed to solve the problem of data association by the active distribution and recognition of RFID tags. Furthermore, RFID tags can be utilized for a communication-free coordination of a robot team and used by human forces to store additional user data such as the number of victims located in a room or an indication of a hazardous area.

The work has been extensively tested on two different robot platforms, a 4WD (four wheeled drive) differentially steered robot for the autonomous team exploration of large office-like arenas, and a tracked robot for climbing 3D obstacles.

ICRA'07 Paper: RFID-Based Exploration for Large Robot Teams - [via] Link

[MIT Thesis] Distributed Method Selection and Dispatching of Contingent, Temporally Flexible Plans

Authors:
Block, Stephen

Abstract:
Many applications of autonomous agents require groups to work in tight coordination. To be dependable, these groups must plan, carry out and adapt their activities in a way that is robust to failure and to uncertainty. Previous work developed contingent, temporally flexible plans. These plans provide robustness to uncertain activity durations, through flexible timing constraints, and robustness to plan failure, through alternate approaches to achieving a task. Robust execution of contingent, temporally flexible plans consists of two phases. First, in the plan extraction phase, the executive chooses between the functionally redundant methods in the plan to select an execution sequence that satisfies the temporal bounds in the plan. Second, in the plan execution phase, the executive dispatches the plan, using the temporal flexibility to schedule activities dynamically. Previous contingent plan execution systems use a centralized architecture in which a single agent conducts planning for the entire group. This can result in a communication bottleneck at the time when plan activities are passed to the other agents for execution, and state information is returned. Likewise, a computation bottleneck may also occur because a single agent conducts all processing. This thesis introduces a robust, distributed executive for temporally flexible plans, called Distributed-Kirk, or D-Kirk. To execute a plan, D-Kirk first distributes the plan between the participating agents, by creating a hierarchical ad-hoc network and by mapping the plan onto this hierarchy. Second, the plan is reformulated using a distributed, parallel algorithm into a form amenable to fast dispatching. Finally, the plan is dispatched in a distributed fashion. We then extend the D-Kirk distributed executive to handle contingent plans. Contingent plans are encoded as Temporal Plan Networks (TPNs), which use a non-deterministic choice operator to compose temporally flexible plan fragments into a nested hierarchy of contingencies. A temporally consistent plan is extracted from the TPN using a distributed, parallel algorithm that exploits the structure of the TPN.At all stages of D-Kirk, the communication load is spread over all agents, thus eliminating the communication bottleneck. In particular, D-Kirk reduces the peak communication complexity of the plan execution phase by a factor of O(A/e'), where e' is the number of edges per node in the dispatchable plan, determined by the branching factor of the input plan, and A is the number of agents involved in executing the plan.In addition, the distributed algorithms employed by D-Kirk reduce the computational load on each agent and provide opportunities for parallel processing, thus increasing efficiency. In particular, D-Kirk reduces the average computational complexity of plan dispatching from O(eN^3) in the centralized case, to typical values of O(eN^2) per node and O(eN^3/A) per agent in the distributed case, where N is the number of nodes in the plan and e is the number of edges per node in the input plan.Both of the above results were confirmed empirically using a C++ implementation of D-Kirk on a set of parameterized input plans. The D-Kirk implementation was also tested in a realistic application where it was used to control a pair of robotic manipulators involved in a cooperative assembly task.

Link

Computer Science and Artificail Intelligence Laboratory Technical Report

Sensitive Manipulation

Link

Abstract:

In this approach, manipulation is mainly guided by tactile feedback as opposed to vision. The traditional approach of a highly precise arm and vision system controlled by a model-based architecture is replaced by one that uses a low mechanical impedance arm with dense tactile sensing and exploration capabilities run by a behavior-based architecture.

The robot OBRERO can come gently in contact, explore, lift, and place the object in a different location. It can also detect slippage and external forces acting on a object while it is held. This tack can be done with very light objects with no fixtures and on slippage surface.

[Robot Perception and Learning] Paper: Auditory Evidence Grids

Eric Martinson, Alan Schultz
Proc. of the International Conference on Intelligent Robots and Systems (IROS) 2006

Abstract – Sound source localization on a mobile robot can be a difficult task due to a variety of problems inherent to a real environment, including robot ego-noise, echoes, and the transient nature of ambient noise. As a result, source localization data are often very noisy and unreliable. In this work, we overcome some of these problems by combining the localization evidence over a variety of robot poses using an evidence grid. The result is a representation that localizes the pertinent objects well over time, can be used to filter poor localization results, and may also be useful for global re-localization from sound localization results.

PDF

Lab Meeting 8 March 2007 (ZhenYu):3D Scene Reconstruction from Reflection Images in a Spherical Mirror

Title : 3D Scene Reconstruction from Reflection Images in a Spherical Mirror
Author : Kanbara, M. Ukita, N. Kidode, M. Yokoya, N.
Conference : ICPR2006

Abstract:

This paper proposes a method for reconstructing a 3D scene structure by using the images reflected in a spherical mirror. In our method, the mirror is moved freely within the field of view of a camera in order to observe a surrounding scene virtually from multiple viewpoints. The observation scheme, therefore, allows us to obtain the wide-angle multiviewpoint images of a wide area. In addition, the following characteristics of this observation enable multi-view stereo with simple calibration of the geometric configuration between the mirror and the camera; (1) the distance and direction from the camera to the mirror can be estimated directly from the position and size of the mirror in the captured image and (2) the directions of detected points from each position of the moving mirror can be also estimated based on reflection on a spherical surface. Some experimental results show the effectiveness of our 3D reconstruction method.

[Link]

News: EXTENDABLE BUMPERS COULD KEEP SMART CONVOYS SAFE

EXTENDABLE BUMPERS COULD KEEP SMART CONVOYS SAFE

Extendable bumpers could be used by convoys of intelligent vehicles to maintain safe distances when normal communications fail, according to researchers, who say their simulations suggest that bumpers that expand pneumatically to touch the car ahead could help keep the convoy together safely in an emergency. Prototype automated control systems that allow intelligent vehicles to travel into tight bunches reduce congestion, save fuel and cut down on accidents, but if communications become disrupted, the situation could be too fast-moving and complex for human drivers, with dangerous consequences. A car equipped with sensors in an extendable bumper would allow the vehicle to determine what the one in front was doing and react quickly by measuring compression of the bumper to track the relative speed and acceleration of the car ahead. Read more: the link.

Lab meeting, March 08, 2007 (Vincent)

My talk will be consisted of the following components :

1. Fit a face through Active Appearance Model (AAM).
2. Get the frontal view of the given face.
3. Use eigenface or fisherface to recognize faces.

Lab Meeting 8 March 2007 (Stanley)

I will briefly introduce the planning and control method:
1. The ND method developed by Javier Minguez and Luis Montano and our modification.
2. The global planning module implementing on ITRI robot.

Lab Meeting 8 March 2007 (Casey): A Multiview Face Identification Model With No Geometric Constraints

Title: A Multiview Face Identification Model With No Geometric Constraints
Author:Jerry Jun Yokono, Sony Intelligence Dynamics Laboratories, Inc., Japan
　　　　　　Tomaso Poggio, M.I.T.
Source: Proceedings of the 7th International Conference on Automatic Face and Gesture Recognition (FGR06)

Abstract:
Face identification systems relying on local descriptors areincreasingly used because of their perceived robustness withrespect to occlusions and to global geometrical deformations.Descriptors of this type – based on a set of oriented Gaussian derivative filters – are used in our identification system. In this paper, we explore a pose-invariant multiview face identification system that does not use explicit geometrical information. The basic idea of the approach is to find discriminant features to describe a face across different views. A boosting procedure is used to select features out of a large feature pool of local features collected from the positive training examples. We describe experiments on well-known, though small, face databases with excellent recognition rate.

Download: [link]

CMU ML talk: Features, kernels, and similarity functions

Speaker: Prof. Avrim Blum, CMU
http://www.cs.cmu.edu/~avrim

Title: Features, kernels, and similarity functions

Date: March 05

Abstract: Given a new learning problem, one of the first things you need to do is figure out what features you want to use. Alternatively, there has been substantial work on kernel functions which provide implicit feature spaces, but then how do you pick the right kernel? In this talk I will survey some theoretical results that can provide some help or at least guidance for these tasks. In particular, I will talk about:

* Algorithms designed to handle large feature spaces when it is expected that only a small number of the features will actually be useful (so you can pile a lot on when you don't know much about the domain).

* Kernel functions. Can theory provide some guidance into selecting or designing a kernel function in terms of natural properties of your domain?

* Combining the above. Can we use kernels to generate explicit features?

[in addition to the survey nature, part of this talk will sneak in some work joint with Nina Balcan]

Link for the videos

There are videos about the result of "Grasp Novel Objects":

http://ai.stanford.edu/~asaxena/learninggrasp/

Videos -> Detailed videos

[Robot Perception and Learning] Lab Meeting Fri, 2 March 2007 : Robotic Grasping of Novel Objects

Ashutosh Saxena, Justin Driemeyer, Justin Kearns, Andrew Y.Ng
Computer Science Department
Stanford University, Stanford, CA 94305

[link]


Abstract:

We consider the problem of grasping novel objects, specifically ones that are being seen for the first time through vision. We present a learning algorithm that
neither requires, nor tries to build, a 3-d model of the object. Instead it predicts,
directly as a function of the images, a point at which to grasp the object. Our algorithm is trained via supervised learning, using synthetic images for the training set. We demonstrate on a robotic manipulation platform that this approach successfully grasps a wide variety of objects, such as wine glasses, duct tape, markers, atranslucentbox, jugs, knife-cutters, cellphones, keys, screwdrivers, staplers, toothbrushes, a thick coil of wire, a strangely shaped power horn, and others, none of which were seen in the training set.

Lab Meeting 2 March 2007 (Chihao): Audio Source detection and tracking - sensor model

Sensor model is an important part of tracking system.
I will show the microphone arrays model and the tracking results.

Lab Meeting 2 March 2007 (Jeff): RFID-Based Exploration for Large Robot Teams

Abstract:
To coordinate a team of robots for exploration is a challenging problem, particularly in largeareas as for example the devastated area after a disaster.This problem can generally be decomposed into task assignment and multi-robot path planning. In this paper, we address both problems jointly. This is possible because we reduce significantly the size of the search space by utilizing RFID tags as coordination points.
The exploration approach consists of two parts: a stand-alone distributed local search and a global monitoring process which can be used to restart the local search in more convenient locations. Our results show that the local exploration works for large robot teams, particularly if there are limited computational resources. Experiments with the global approach showed that the number of conflicts can be reduced, and that the global coordination mechanism increases significantly the explored area.

Link:
http://www.informatik.uni-freiburg.de/~kleiner/pubs/ziparo_et_al_icra07-abstract.html

Lab Meeting 2 March 2007 (Eric): Projector and Camera Syatem - 3D Model Construction

I will show the result based on the point to surface scan-matching and discuss how it works.

News: DIY Bluetooth Accelerometer

College Of Computing at Georgia Tech have a great DIY Bluetooth accelerometer with source and schematic included.

Overview:

This is a small wireless sensor platform providing a bluetooth SPP link to three axes of accelerometer data. The accelerometers are sampled by a PIC microcontroller (onboard ADC) at roughly 100Hz (rate can be changed via firmware). Data from the ADC conversion is sent to a remote computer using the PIC's UART in conjunction with a drop-in bluetooth serial part. Even with two dual-axis accelerometers onboard there are up to 17 free I/O lines and two additional ADC channels depending on the device configuration. Three sockets provide access to all PIC signals. Over-the-air programming allows for easy firmware updates and rapid prototyping without the need to have a PIC programmer or special cable. Schematics, parts lists, and firmware sources are available online.

Feature List:

• Bluetooth serial port profile (SPP) for standardized interface
• 3-Axis accelerometer data, 3.9mg resolution (ADXL202JE)
• Reprogrammable PIC (16LF873-04I)
• 17 free I/O lines
• In-Circuit programming connector
• Over-the-air programming via bootloader
• Battery life ~60hrs on 840mAh 3.7v battery, full TX mode
• 13mA @ 3.7v TX mode, 3mA @ 3.7v standby
• Very Simple Firmware
• 35mm X 35mm X 5mm
  

Contextual Computing Group: Bluetooth Accelerometer - Link

CMU VASC talk: Clustering and Classification via Lossy Data Compression

Clustering and Classification via Lossy Data Compression

Yi Ma, UIUC
Monday, Feb 26

For many problems in computer vision, image processing, and pattern recognition, we need to process and analyze massive amount of high-dimensional mixed data such as images and gene expression data. By mixed data, we mean that the given data set consists of multiple heterogeneous subsets (which have different geometric or statistical characteristics) but each subset can be more easily modeled or representedthan the whole data set together.

In this talk, we address two fundamental questions: .How to cluster and classify such high-dimensional mixed data?. We contend that both the (unsupervised) clustering and (supervised) classification problems can be cast as a lossy data compression problem and solved efficiently within a unified mathematical framework. In theory, this approach offers some distinguished advantages over conventional methods for clustering and classification, especially in dealing with several difficult issues that often arise in practice: regularization of degenerate distributions, selection of models with different complexities, and rejection of outliers.

Our work establishes a strong connection between information theory, especially the rate-distortion theory, with data clustering and classification, and it leads to extremely simple but effective algorithms. We will demonstrate the success of these algorithms in a few popular but difficult problems, including but not limited to natural image segmentation, microarray data clustering, handwritten digits and face
recognition.

CMU talk: Interacting Physically with Robots and Virtually on the Global Digital Campus

SCS DISTINGUISHED LECTURE SERIES
Yuichiro Anzai
President, Keio University

Abstract:
The lecture presents two topics: one for the research on human-robot interaction conducted in the Anzai-Imai laboratory and the other for the activities of the Research Institute for Digital Media and Content, both at Keio University.
Our research on human-robot interaction, embarked upon in 1991, is concerned with designing technologies that facilitate the smooth interaction of humans with robots. We initially started by designing software and hardware systems that support human-robot interaction, and then moved forward, with Michita Imai and others, to designing robots that can smoothly interact with humans. In some cases we conducted behavioral experiments to find out how a human behaves in an interaction with a robot, and fed the results back to engineering. The first part of the lecture provides a summary of the efforts at our lab during these fifteen years.
The second part of the lecture will focus on the activities of the Research Institute for Digital Media and Content, established in 2004. One of its goals is to use various technologies to extend the reach of our physical campus so that students and faculty members can distribute their academic knowledge to a global audience, interact with people around the world, and have convenient access to globally shared knowledge. We have already set up what we call Global Digital Studios in Tokyo, Seoul, Beijing, Cambridge (UK) and San Francisco, with others scheduled to open in New York and some other locations. Twenty-four higher learning institutions in twelve South-East Asian countries are also tied to this network via satellite Internet. The Studios and sites can be connected online at any time, and are used for many different purposes: the network can be regarded as an early version of our Global Digital Campus. The second part of the lecture gives a glimpse of this effort at Keio University.

Bio:
Born in 1946 in Tokyo, Yuichiro Anzai received his Ph.D. in engineering from Keio University in 1974. After serving at Keio as an assistant professor until 1985, he joined the faculty of Hokkaido University as an associate professor in behavioral science. In 1988 he returned to Keio as a professor in electrical engineering, and became the dean of the Faculty of Science and Technology in 1993. He worked on the reform of the undergraduate departments and graduate programs for more than seven years, and launched new educational and research programs with an innovative structure. Since 2001, he has served as president of Keio University, the oldest modern institution of higher learning in Japan (http://www.keio.ac.jp/index-en.htm) that will celebrate its 150th anniversary in 2008, as well as a professor in the Department of Information and Computer Science and the School of Open and Environmental Systems. At present, much of his time is devoted to driving forward the commemorative fundraising campaign and associated programs (http://keio150.jp/english).
Professor Anzai was a post-doc in the Departments of Psychology and Computer Science, and a visiting assistant professor in the Department of Psychology, Carnegie-Mellon University, in 1976-78, and in 1981-82 respectively. He was also a visiting professor at the Center for Medical Education, McGill University, in 1990. His fields of research include cognitive science and computer science, particularly cognitive processes in learning and problem solving, and human-robot-computer interaction. He has published about 20 books, single- and co-authored, and more than 120 technical papers in those fields. For public service, he is serving as president of the Information Processing Society of Japan, as president of the Japan Association of Private Universities and Colleges, as a member of the Science Council of Japan, and as a member of the Central Council for Education.

CMU talk: The Structure and Acquisition of Semantic Knowledge

The Structure and Acquisition of Semantic Knowledge

Charles Kemp
Department of Brain & Cognitive Sciences
Massachusetts Institute of Technology

Thursday, February 22, 2007

Humans regularly make inferences that go beyond the data they have seen. Two questions immediately arise: what is the knowledge that supports these inferences, and how is this knowledge acquired? I will present a hierarchical Bayesian approach to inductive reasoning that addresses both questions. When making inferences about the distribution of a novel property, people draw on rich semantic knowledge that can often be captured using structured representations of the relationships between the entities in a domain. For instance, given that gazelles have T4 cells and carry E. spirus bacteria, taxonomic relations are useful for predicting which other animals are likely to have T4 cells, but predator-prey relations are more useful when reasoning about the distribution of E. spirus bacteria. I will show that our formal framework provides close quantitative fits to human inferences about several kinds of properties when supplied with appropriate knowledge representations for each task. Different inductive tasks often draw on different kinds of knowledge which are best captured by qualitatively different kinds of representations. For instance, anatomical features of biological species are best captured by a taxonomic tree, political views are best captured by a linear spectrum, and friendship relations are best captured by a set of discrete cliques. Our hierarchical framework helps to explain how humans can discover the best kind of representation for a given inductive context.

CMU talk: Adaptive Online Allocation Mechanisms for Single-Valued Domains

Adaptive Online Allocation Mechanisms for Single-Valued Domains

David C. Parkes, Harvard University

Abstract: Mechanism design studies the problem of designing protocols that will implement desirable outcomes in multi-agent systems with self-interest and private information. Many interesting domains are inherently dynamic with uncertainty about both supply and demand; e.g., selling seats on an airplane, adverts on a search engine, computational resources. The classic Vickrey-Clarke-Groves mechanism extends to dynamic environments but is non-adaptive and much less robust than when used offline. Our interest in this talk is in the design of adaptive, online allocation mechanisms, that are able to leverage a probabilistic (perhaps incorrect) model of the environment. We focus on single-valued domains in which agents are indifferent across one of a set of equivalent allocations. A truthful, online stochastic optimization algorithm coupled with historical sampling and an ``ironing" procedure is presented, along with examples to show that the optimal policy is generally not truthfully implementable. Simulation analysis illustrates the cost of truthfulness in application to selling a computational resource.

Bio: David C. Parkes is the John L. Loeb Associate Professor of the Natural Sciences and Associate Professor of Computer Science at Harvard University. He received his Ph.D. degree in Computer and Information Science from the University of Pennsylvania in 2001, and an M.Eng. (First class) in Engineering and Computing Science from Oxford University in 1995. He was awarded the prestigious NSF CAREER Award in 2002, an IBM Faculty Partnership Award in 2002 and 2003, and the Alfred P. Sloan Fellowship in 2005. Parkes has published extensively on topics related to electronic markets, computational mechanism design, auction theory, and multi-agent systems. He serves on the editorial board of the Journal of Artificial Intelligence Research and the Electronic Commerce Research Journal, and has served on the Program Committee of a number of leading conferences in artificial intelligence, multiagent systems and electronic commerce, including ACM-EC, AAAI, IJCAI, UAI and AAMAS. Parkes is the co program-chair of the ACM Conference on Electronic commerce, 2007 and the Int. Conf. on Autonomous Agents and Multiagent systems, 2008.

http://www.eecs.harvard.edu/~parkes