Seminars

Seminars From Year 2012

View seminars by year:     2003    2004    2005    2006    2007    2008    2009    2010    2011    2012    2013    2014    All    

All seminars will be held in Wu & Chen Auditorium, Levine Hall (3330 Walnut Street) unless otherwise indicated.

  • Talk: Spring 2012 GRASP Seminar: Raffaello D'Andrea, Swiss Federal Institute of Technology, Zurich, "Robustness by Necessity: Zero-Downtime Demos, Competitions, and Live Performances"
    Date: Friday, January 20, 2012 - 11am to 12pm
    Presenters: Raffaello D'Andrea
    A key bottleneck preventing the widespread deployment of robotic and autonomous systems is robustness. Unfortunately, the university research environment does little to address this problem: The emphasis is almost exclusively on new algorithms and capabilities, with insufficient attention paid to their inherent robustness. In this talk I discuss several ways in which we have been able to naturally incorporate robustness in our research: by putting ourselves in positions where the cost of failure is high.
  • Talk: Spring 2012 GRASP Seminar - Jingyi Yu, University of Delaware, "Beyond Perspective Cameras: Multi-perspective Imaging, Reconstruction, Rendering and Projection"
    Date: Friday, January 27, 2012 - 11am to 12pm
    Presenters: Jingyi Yu
    A perspective image represents the spatial relationships of objects in a scene as they appear from a single viewpoint. In contrast, a multi-perspective image combines what is seen from several viewpoints into a single image. Despite their incongruity of view, effective multi-perspective images are able to preserve spatial coherence and can depict, within a single context, details of a scene that are simultaneously inaccessible from a single view. In this talk, I will provide a complete framework for using multi-perspective imaging models in computer vision and graphics. Our multi-perspective framework consists of four key components: acquisition, reconstruction, rendering, and display. A multi-perspective camera captures a scene from multiple viewpoints in a single image. From the input image, intelligent software can recover 3D scene geometry using multi-perspective stereo matching algorithms or via shape-from-distortion approaches. Our solutions are particularly useful for reconstructing specular (reflective and refractive) surfaces such as dynamic 3D fluid surfaces which can also be viewed as general multi-perspective cameras. The recovered geometry, along with lighting and surface reflectance, can then be loaded into a new multi-perspective graphics pipeline for real-time rendering. Finally, we can visualize the rendering results on a special multi-perspective display that combines a single consumer projector and specially-shaped mirrors/lenses. Such displays will offer an unprecedented level of flexibility in terms of aspect ratio, size, field of view, etc.
  • Talk: Spring 2012 GRASP/HMS Seminar: Aaron Hertzmann, University of Toronto, "Feature-Based Locomotion Controllers for Physically-Simulated Characters"
    Date: Friday, February 3, 2012 - 11am to 12pm
    Presenters: Aaron Hertzmann
    Understanding the control forces that drive humans and animals is fundamental to describing their movement. Although physics-based methods hold promise for creating animation, they have long been considered too difficult to design and control. Likewise, recent results in computer vision suggest how physical models, if developed, could be important to human pose tracking. I will describe the main problems of human motion modeling. I will then present a new approach to control of physics-based characters based on high-level features of human movement. These controllers provide unprecedented flexibility and generality in real-time character control: they are capture many natural properties of human movement; they can be easily modified and applied to new characters; and they can handle a variety of different terrains and tasks, all within a single control strategy. Until very recently, even making a controller walk without falling down was extraordinarily difficult. This is no longer the case. Our work, together with other recent results in this area, suggests that we are now ready to make great strides in locomotion.
  • Talk: Spring 2012 GRASP Seminar - Satyandra K. Gupta, University of Maryland, "Planning for Autonomous Robotic Operations in Physically Challenging Environments"
    Date: Friday, February 10, 2012 - 11am to 12pm
    Presenters: Satyandra K. Gupta
    Autonomous robotic operations require robots to be able to automatically generate plans. Physically challenging environments require robots to be able to negotiate around dynamically moving objects, cope with significant uncertainties in the outcome of action execution, sensor limitations, and the presence of intelligent adversaries. This seminar will cover the following four topics. First, I will describe a planning architecture that integrates task planning, behavior selection, and trajectory planning in a seamless manner to successfully handle physically challenging environments. This approach provides the right balance between deliberative and reactive planning during the execution of complex tasks in a dynamic uncertain environment. Second, I will describe our work in the area of physically accurate computationally efficient simulations to enable physics-aware planning. Third, I will describe computational synthesis techniques for automatically generating sophisticated reactive behaviors. This synthesis approach automatically generates an initial version of an action selection policy and then gradually refines it by detecting and fixing its shortcomings. Finally, I will describe an approach for integrating game tree search concepts within our planning framework to manage risks. The following applications will be used to illustrate the approach: (1) guarding of a valuable asset by autonomous unmanned sea surface vehicles, (2) supply mission on a rugged terrain by unmanned ground vehicles, and (3) assembly of micro particles in a fluidic medium using holographic optical tweezers.
  • Talk: GRASP Special Seminar - Andrzej Banaszuk, Suresh Kannan & Alberto Speranzon, UTRC, "Autonomous and Intelligent Systems at United Technologies Research Center"
    Date: Wednesday, February 15, 2012 - 3pm to 4pm
    Presenters: Andrzej Banaszuk, Suresh Kannan & Alberto Speranzon
    Alternate Location: Levine 307 (3330 Walnut Street)
    We will present UTRC’s research initiative in Autonomous and Intelligent Systems with an emphasis on complex human/machine intelligent systems including unmanned rotorcraft. The research, conducted by a diverse team of researchers in dynamical systems, control, applied mathematics, computer vision, and embedded systems (in partnership with several leading universities including CMU, MIT, Caltech, and BU) includes: Real-time algorithms for dynamic collision avoidance in an obstacle-rich environment using probabilistic roadmaps.Multi-vehicle missions including efficient search algorithms based on ergodic theory methods.Multi-vehicle navigation with imperfect and intermittent sensors in GPS degraded environments.Intelligent system design methodology including architectures for autonomy, formal verification, and stochastic planners. We will present research problems of interest to UTRC and discuss existing and future career opportunities in the broad area of robotics, with particular focus on perception, machine intelligence, and human-machine interaction.
  • Talk: Spring 2012 GRASP Seminar: Christopher Geyer, IRobot, "Heads-up, hands-free operation of UGVs through Embedded Computer Vision"
    Date: Friday, February 17, 2012 - 11am to 12pm
    Presenters: Christopher Geyer
    Today, most commercially available UGVs use teleoperation for control.  Under teleoperation, users’ hands are tied up holding a hand-held controller to operate the UGV, and their attention is focused on what the robot is doing.  In this talk, I will describe iRobot’s work on an alternative to teleoperation called Heads-up, Hands-free Operation, which allows an operator to control a UGV using operator following behaviors and a gesture interface.  We explore whether Heads-up, Hands-free Operation is an improvement over teleoperation.  In a study of 30 participants, we found that when operators used these modes of interaction, they performed missions faster, they could recall their surroundings better, and they had a lower cognitive load than they did when they teleoperated the robot. Heads-up, Hands-free Operation is enabled through the use of embedded computer vision technologies.
  • Talk: Spring 2012 GRASP Seminar: Jonathan Clark, FAMU/FSU College of Engineering, "Design of Dynamic Multi-Modal Legged Locomotors"
    Date: Friday, February 24, 2012 - 11am to 12pm
    Presenters: Jonathan Clark
    Despite substantial progress in robotic systems over the past couple of decades, animals continue to set the standard for locomotion prowess in unknown and cluttered environments, whether they are natural or man-made.   Finely tuned legged systems that explicitly exploit their body’s natural dynamics have begun to rival specific performance criteria, such as speed over smooth terrain, of the most accomplished biological systems.  Our ability, however to mimic their ability to operate in a multi-modal fashion (running, climbing, leaping, flying, and/or crawling, etc.) severely limits our locomotive ability in complex environments.  In this talk I will address some of the issues associated with the design of these systems, including identifying and anchoring appropriate low-dimensional dynamic models and efficient exploitation of available onboard power for fast and stable locomotion. 
  • Talk: Spring 2012 GRASP Seminar - Emily B. Fox, University of Pennsylvania, "Bayesian Nonparametric Methods for Complex Dynamical Phenomena"
    Date: Friday, March 2, 2012 - 11am to 12pm
    Presenters: Emily B. Fox
    Markov switching processes, such as hidden Markov models (HMMs) and switching linear dynamical systems (SLDSs), are often used to describe rich classes of dynamical phenomena.  They describe complex temporal behavior via repeated returns to a set of simpler models: imagine, for example, a person alternating between walking, running and jumping behaviors, or a stock index switching between regimes of high and low volatility. Traditional modeling approaches for Markov switching processes typically assume a fixed, pre-specified number of dynamical models.  Here, in contrast, I develop Bayesian nonparametric approaches that define priors on an unbounded number of potential Markov models.  Using stochastic processes including the beta and Dirichlet process, I develop methods that allow the data to define the complexity of inferred classes of models, while permitting efficient computational algorithms for inference.  The new methodology also has generalizations for modeling and discovery of dynamic structure shared by multiple related time series. Interleaved throughout the talk are results from studies of the NIST speaker diarization database, stochastic volatility of a stock index, the dances of honeybees, and human motion capture videos.
  • Talk: GRASP Special Seminar - Kris Bhaskar VP of Systems & Dr. Eliezer Rosengaus Senior Research Scientist, CTO office, KLA-Tencor Corp, "KLA-Tencor"
    Date: Thursday, March 22, 2012 - 12pm to 1pm
    Presenters: Kris Bhaskar VP of Systems & Dr. Eliezer Rosengaus Senior Research Scientist
    A KLA-Tencor perspective on the challenges and solutions pertaining to semi-conductor manufacturing and yield management using image processing and computer vision technologies. In addition, the presentation will also cover ideas for potential research collaboration in the future between U-Penn and KLA-Tencor in a variety of related fields.
  • Talk: Spring 2012 GRASP Seminar - Serge Belongie, University of California, San Diego, "Visual Recognition with Humans in the Loop"
    Date: Friday, March 23, 2012 - 11am to 12pm
    Presenters: Serge Belongie
    We present an interactive, hybrid human-computer method for object classification. The method applies to classes of problems that are difficult for most people, but are recognizable by people with the appropriate expertise (e.g., animal species or airplane model recognition). The classification method can be seen as a visual version of the 20 questions game, where questions based on simple visual attributes are posed interactively. The goal is to identify the true class while minimizing the number of questions asked, using the visual content of the image. Incorporating user input drives up recognition accuracy to levels that are good enough for practical applications; at the same time, computer vision reduces the amount of human interaction required. The resulting hybrid system is able to handle difficult, large multi-class problems with tightly-related categories. We introduce a general framework for incorporating almost any off-the-shelf multi-class object recognition algorithm into the visual 20 questions game, and provide methodologies to account for imperfect user responses and unreliable computer vision algorithms. We evaluate the accuracy and computational properties of different computer vision algorithms and the effects of noisy user responses on a dataset of 200 bird species and on the Animals With Attributes dataset. Our results demonstrate the effectiveness and practicality of the hybrid human-computer classification paradigm.This work is part of the Visipedia project, in collaboration with Steve Branson, Catherine Wah, Florian Schroff, Boris Babenko, Peter Welinder and Pietro Perona.
  • Talk: Spring 2012 GRASP Seminar: James McLurkin, Rice University, "Distributed Algorithms for Robot Recovery, Angular Coordinate Systems, and Low-Cost Robots: An Overview of the Rice Multi-Robot Systems Lab"
    Date: Friday, March 30, 2012 - 11am to 12pm
    Presenters: James McLurkin
    In this talk we present results from three different projects: 1. A distributed recovery algorithm to extract a multi-robot system from complex environments.  The goal is to maintain network connectivity while allowing efficient recovery.  Our approach uses a maximal-leaf spanning tree as a communication and navigation backbone, and routes robots along this tree to the goal.  Simulation and experimental results demonstrate the efficacy of this approach.  2. Angular coordinate systems can provide robots with useful network geometry from very low-cost hardware. We introduce "scale-free coordinates" as a coordinate system of intermediate power and design complexity.  We show that it can estimate low-quality network geometry, but can still be used to build a useful motion controller with interesting limitations.  3. We introduce the "r-one" robot, a low-cost design suitable for research, education, and outreach.  We provides tales of joy and disaster from using 90 of these platforms for our research, a freshman engineering systems course, and graduate robotics lab.
  • Talk: GRASP Special Seminar - Roberto Tron, Johns Hopkins University, "Consensus Algorithms on Manifolds and Localization of Camera Networks"
    Date: Monday, April 2, 2012 - 1pm to 2pm
    Presenters: Roberto Tron
    Alternate Location: Levine 315 (3330 Walnut Street)
    Consensus algorithms are a popular choice for computing averages and other aggregate functions in ad-hoc wireless sensor networks. However, existing work mostly addresses the case where the measurements lie in a Euclidean space. In the first part of the talk we will present Riemannian consensus, a natural extension of consensus algorithms to Riemannian manifolds. We will discuss its convergence properties and their dependence on various factors, such as choice of the distance function, network connectivity, geometric configuration of the measurements and curvature of the manifold. In the second part of the talk we will focus on the problem of distributed 3-D camera network localization, and show how ideas and analysis techniques from Riemannian consensus can be extended and applied to this problem.
  • Talk: GRASP Special Seminar - Brandon Basso, University of California, Berkeley, "Learned Vehicle Routing"
    Date: Tuesday, April 10, 2012 - 10am to 11am
    Presenters: Brandon Basso
    Alternate Location: Levine 512 (3330 Walnut Street)
    Autonomous decision making is a broad field that encapsulates many topics germane to autonomous systems. The vehicle routing problem (VRP) is a particularly well-studied decision problem in which the goal is to find an optimal allocation of agents to tasks while respecting spatial, temporal, and capacity constraints. Air traffic control, postal service package delivery, and many other similar applications can generally be categorized as multi-agent vehicle routing. Traditional solution approaches have a difficulty in coping with such broad problems, characterized by vast state spaces, evolving constraints, unstructured environments, and ambiguous state transition dynamics. Departing from traditional optimization and search-based approaches, this work seeks to pose vehicle-routing-based problems in a machine learning context. A representation scheme is presented that scales independently of the physical problem size. The abstracted cost-based state captures an agent’s ability to perform a certain task, transferring complexity from constraints into the state itself. In order to capture temporal constraints such as deadlines without increasing complexity, a semi-Markov model is proposed and analyzed relative to a traditional Markov model. Well-known iterative learning algorithms based on dynamic programming readily solve for optimal policies and can be compared to known VRP solutions. Results borne out in simulation demonstrate the benefits of modeling high-level decision problems such as VRPs in a learning framework.
  • Talk: Spring 2012 GRASP Seminar: Chad Jenkins, Brown University, "rosbridge: Towards a World Wide Web for Robotics"
    Date: Friday, April 13, 2012 - 11am to 12pm
    Presenters: Chad Jenkins
    Our work aims to make robot technology a seamless part of the larger World Wide Web, such as through applications-layer robotics protocols. We posit that the convergence of robotics with Internet and Web technologies will lead to a thriving robotics ecosystem with greater levels of reproducibility, interoperability, and accessibility. Broader populations of users and applications developers will be able to use and create "apps" that address various needs across society, while building on continuing advances in robotics research and development. Currently, other than online videos, the impact of the great achievements and capabilities in robotics remains relatively limited to robotics research labs. Robotics has made great strides in producing a variety of capable and affordable off-the-shelf platforms comprised of physically capable hardware, richer perception of unstructured environments, and general purpose robot middleware. However, these robots lack a common and general format for exchanging information that both limits interoperability between systems and accessibility of interfaces for developing applications for broader human use. In this talk, I will cover our recent work with rosbridge as a lightweight applications-layer protocol for robotics. Assuming only the JSON format and network sockets, rosbridge was originally intended to enable any network process (independent of any specific operating system or build environment) to access, share topic messages with, and call services provided by a Robot Operating System (ROS) run-time environment. More generally, rosbridge is a step towards a larger robotics World Wide Web, where protocols provide generic messaging and data exchange between robots and robot middleware provides server-side functionality (analogous to a web server). Using rosbridge, I describe our robot web applications implemented purely through JavaScript/HTML for web-scale robot learning and a PR2 Remote Lab. Such applications demonstrate "no-install" interfaces for reaching broader populations of users as well as platforms for common decentralized experimentation. I will also discuss examples of improved interoperability in robotics projects enabled by rosbridge.
  • Talk: GRASP Special Seminar - Geoffrey Barrows, Founder of Centeye, "Bio-Inspired Vision for Micro Air Vehicles"
    Date: Friday, April 13, 2012 - 2pm to 3pm
    Presenters: Geoffrey Barrows
    Alternate Location: Wu and Chen Auditorium (Levine 101)
    Flying insects have long been accepted as a good model for developing vision systems for micro air vehicles. In this talk, I will relate my own experiences implementing bio-inspired vision and flight control to provide small scale navigation e.g. obstacle avoidance and stability to air vehicles of this size. This talk will emphasize practical lessons learned that the audience should be able to take away and apply to their own research. These lessons include the benefits of using a “vertically integrated” approach, the pitfalls of using more pixels or features than what is needed for a task, and that one does not necessarily need to know the 6DOF Cartesian state of an air vehicle in order to control it.
  • Talk: GRASP/PRECISE Seminar: Eric Feron, Georgia Institute of Technology, "Proof-Carrying Auto-Coded Control Software"
    Date: Friday, April 13, 2012 - 3pm to 4pm
    Presenters: Eric Feron
    Alternate Location: Wu and Chen Auditorium (Levine 101)
    Proof-carrying code has been in existence since Necula and Lee coined the term in 1996. This talk brings forward the details of the development of proof-carrying code from control system specifications. The motivation for this work is the safety-critical nature of many control applications such as aeronautics, robot-assisted surgery, and ground transportation. Several challenges must be addressed during this effort, including: The formal representation of control-theoretic proofs; the migration and representation of these proofs across different layers of software implementation; and the design of a back-end to verify the claimed software properties. The expected payoff from these efforts is to include more semantics in the output of computer-aided control system design environments and to influence the software certification processes currently in use for transportation and health applications.
  • Talk: Spring 2012 GRASP Seminar - Russell H. Taylor, Johns Hopkins University, "A Microsurgery Assistant System for Retinal Surgery"
    Date: Friday, April 20, 2012 - 11am to 12pm
    Presenters: Russell H. Taylor
    This talk will discuss ongoing NIH-funded research at Johns Hopkins University and Carnegie-Mellon University to develop technology and systems addressing fundamental limitations in current microsurgical practice, using vitreoretinal surgery as our focus.   Vitreoretinal surgery is the most technically demanding ophthalmologic discipline and addresses prevalent sight-threatening conditions in areas of growing need.  At the center of our planned approach is a “surgical workstation” system interfaced to a stereo visualization subsystem and a family of novel sensors, instruments, and robotic devices.  The capabilities of these components individually address important limitations of current practice; together they provide a modular, synergistic, and extendable system that enables computer-interfaced technology and information processing to work in partnership with surgeons to improve clinical care and enable novel therapeutic approaches.  The talk will also talk briefly about other medical robotics research at Johns Hopkins University to develop systems that combine innovative algorithms, robotic devices, imaging systems, sensors, and human-machine interfaces to work cooperatively with surgeons in the planning and execution of surgery and other interventional procedures.  Here, we will pay special attention to joint projects between JHU and Intuitive Surgical, including our efforts to develop an open-source “Surgical Assistant Workstation” software environment to promote research and technology transfer activities.
  • Talk: Spring 2012 GRASP Seminar: Kristen Grauman, University of Texas, "Focusing Human Attention on the “Right” Visual Data"
    Date: Friday, April 27, 2012 - 11am to 12pm
    Presenters: Kristen Grauman
    Alternate Location: Berger Auditorium (Skirkanich Hall, Room 13)
    Widespread visual sensors and unprecedented connectivity have left us awash with visual data---from online photo collections, home videos, news footage, medical images, or surveillance feeds.  Which images and videos among them warrant human attention?  I present two problem settings in which this question is critical: supervised learning of object categories, and unsupervised video summarization.  In the first setting, the challenge is to sift through candidate training images and select those that, if labeled by a human, would be most informative to the recognition system.  To address this challenge, we introduce a novel large-scale active learning algorithm that efficiently indexes millions of unlabeled instances according to their informativeness.  We use it to deploy a “live learning” system that actively requests crowd-sourced annotations on images crawled from the Web, yielding state-of-the-art accuracy on the PASCAL object detection benchmark with minimal human intervention. In the second setting, the challenge is to sift through a long-running video and select only the essential parts needed to summarize it for a human viewer.  Unlike traditional keyframe selection techniques, we propose an object-driven approach that predicts the impact each object has on generating the “story” of the video.  Using novel importance cues indicative of importance in an egocentric camera’s view, our approach turns hours of video into a compact storyboard summary that a human can interpret in just seconds. Both domains demonstrate the importance of isolating the key visual data that deserves human attention, and suggest exciting new applications for large-scale visual analysis. This talk describes work with Yong Jae Lee, Sudheendra Vijayanarasimhan, and Prateek Jain. 
  • Talk: GRASP Special Seminar - Michelle Johnson, Marquette University, "Insights into Motor and Brain Changes after Robot-Assisted NeuroRehabilitation"
    Date: Monday, May 21, 2012 - 11am to 12pm
    Presenters: Michelle Johnson
    Alternate Location: Heilmeier Hall (100 Towne)
    Robot-assisted therapy is on the cutting edge of stroke rehabilitation and is a therapy method that promises to improve the lives of persons with disabilities due to stroke.  Preliminary studies using robotic tools provide mixed evidence for their effectiveness and reveal limitations.  There is a need to study the stroke recovery process and to understand how to optimize robot-assisted therapies in order to enhance patient rehabilitation and improve functional outcomes.  Imaging techniques such as functional Magnetic Resonance Imaging (fMRI) and Diffusion Tensor Imaging (DTI) can assist us in understanding the stroke recovery process, in determining who may benefit from robot-assisted training, and in defining how training induced functional cortical changes occur after robot training.  We are conducting an interventional study plus control to assess the effectiveness of robot-assisted practice of tasks with skilled functional tasks.  We assess the ability of active-assisted reaching and grasping training to effect immediate gains and long-term functional improvements in unilateral and bilateral motor control.  Using fMRI and DTI, we associate changes seen in motor impairment levels and functional task performance levels with white matter injuries and connectivity and changes in oxygen utilization in the motor cortex as well as other areas of the brain.  This talk will present insights from case studies with stroke survivors that help us understand motor and brain changes after stroke.
  • Talk: GRASP Special Seminar - Hawkeye King, University of Washington, "Telerobotic Surgery Research With the Raven Surgical System"
    Date: Friday, June 8, 2012 - 11am to 12pm
    Presenters: Hawkeye King
    Alternate Location: Wu and Chen Auditorium (Levine 101)
    Surgical robots create computer mediated, telepresence interaction with a remote operative area. This opens the door to numerous capabilities including remote teleoperation, and computer guidance via haptic tactile overlays, i.e., virtual fixtures. Also, since this is a very new (<2 decades) field, there are few established standards in telesurgery with divergent technologies. This talk presents my ongoing research at the UW BioRobotics Laboratory in telepresence surgery. I will discuss past experiments in remote telesurgery and telerobotic interoperability. Also, I will give details of the recently completed Raven II research surgical robot which is now in use by research groups around the U.S. and describe some of the research being done with the system at the UW.
  • Talk: GRASP Special Seminar - Min Sun, University of Michigan, "Toward Efficient and Robust Human Pose Estimation"
    Date: Thursday, June 14, 2012 - 10am to 11am
    Presenters: Min Sun
    Alternate Location: Levine 512 (3330 Walnut Street)
    Robust human pose estimation is a challenging problem in computer vision in that body part configurations are often subject to severe deformations and occlusions. Moreover, efficient pose estimation is often a desirable requirement in many applications. The trade-off between accuracy and efficiency has been explored in a large number of approaches. On the one hand, models with simple representations (like tree or star models) can be efficiently applied in pose estimation problems. However, these models are often prone to body part misclassification errors. On the other hand, models with rich representations (i.e., loopy graphical models) are theoretically more robust, but their inference complexity may increase dramatically. In this talk, we present an efficient and exact inference algorithm based on branch-and-bound to solve the human pose estimation problem on loopy graphical models. We show that our method is empirically much faster (about 74 times) than the state-of-the-art exact inference algorithm [Sontag et al. UAI'08]. By extending a state-of-the-art tree model [Sapp et al. ECCV'10] to a loopy graphical model, we show that the estimation accuracy improves for most of the body parts (especially lower arms) on popular datasets such as Buffy [Ferrari et al. CVPR'08] and Stickmen [Eichner and Ferrari BMVC'09] datasets. Our method can also be used to exactly solve most of the inference problems of Stretchable Models [Sapp et al. CVPR'11] on video sequences (which contains a few hundreds of variables) in just a few minutes. Finally, we show that the novel inference algorithm can potentially be used to solve human behavior understanding and biological computation problems.
  • Talk: GRASP Special Seminar - Tal Hassner, The Open University of Israel, "Subspaces, SIFTs, and Scale Invariance"
    Date: Friday, June 15, 2012 - 11am to 12pm
    Presenters: Tal Hassner
    Alternate Location: Levine 315 (3330 Walnut Street)
    Scale invariant feature detectors often find stable scales in only a few image pixels. Consequently, methods for feature matching typically choose one of two extreme options: matching a sparse set of scale invariant features, or dense matching using arbitrary scales. In this talk we turn our attention to the overwhelming majority of pixels, those where stable scales are not found by standard techniques. We ask, is scale-selection necessary for these pixels, when dense, scale-invariant matching is required and if so, how can it be achieved? We will show the following: (i) Features computed over different scales, even in low-contrast areas, can be different; selecting a single scale, arbitrarily or otherwise, may lead to poor matches when the images have different scales. (ii) Representing each pixel as a set of SIFTs, extracted at multiple scales, allows for far better matches than single-scale descriptors, but at a computational price. Finally, (iii) each such set may be accurately represented by a low-dimensional, linear subspace. A subspace-to-point mapping may further be used to produce a novel descriptor representation, the Scale-Less SIFT (SLS), as an alternative to single-scale descriptors. The talk will cover these contributions, as well as review our related earlier work on subspace-to-point mappings.
  • Talk: Joint GRASP and Biomedical Image Computing and Informatics Seminar - Nikos Paragios, Ecole Centrale de Paris, "Inverse Problem Modeling and Inference through Graphical Models and Discrete Optimization in Biomedical image analysis"
    Date: Thursday, June 21, 2012 - 2pm to 3pm
    Presenters: Nikos Paragios
    Alternate Location: Wu and Chen Auditorium (Levine 101)
    Inverse modeling aims at recovering the set of parameters of a parametric model such that the resulting instance optimally explains the observations/measurements. Such a formulation is often addressed as an optimization problem of an appropriately defined objective function that is in most of the cases is an ill-posed, highly non-linear and highly non-convex problem. In this seminar, we will investigate the use of graphical models (low and higher order rank) to address such inference and present efficient optimization algorithms that can produce either computational efficient near-optimal solutions or optimal ones through tight relaxations and dual decomposition even for higher order models. The domain of medical imaging and computer vision (segmentation/registration/matching & beyond) will be used to demonstrate the extreme potentials of such modeling and inference methods.
  • Talk: GRASP Special Seminar - Xiaoweii Zhou, Hong Kong University of Science and Technology, "Moving Object Detection by Detecting Contiguous Outliers in the Low-Rank Representation"
    Date: Monday, June 25, 2012 - 11am to 12pm
    Presenters: Xiaoweii Zhou
    Alternate Location: Levine 512 (3330 Walnut Street)
    Object detection is a fundamental step for automated video analysis in many vision applications. Object detection in a video is usually performed by object detectors or background subtraction techniques. Often, an object detector requires manually labeled examples to train a binary classifier, while background subtraction needs a training sequence that contains no objects to build a background model. To automate the analysis, object detection without a separate training phase becomes a critical task. People have tried to tackle this task by using motion information. But existing motion-based methods are usually limited when coping with complex scenarios such as nonrigid motion and dynamic background. In this paper, we show that above challenges can be addressed in a unified framework named DEtecting Contiguous Outliers in the LOw-rank Representation (DECOLOR). This formulation integrates object detection and background learning into a single process of optimization, which can be solved by an alternating algorithm efficiently. We explain the relations between DECOLOR and other sparsity-based methods. Experiments on both simulated data and real sequences demonstrate that DECOLOR outperforms the state-of-the-art approaches and it can work effectively on a wide range of complex scenarios.
  • Talk: GRASP REU Site Oral Presentations - Summer 2012
    Date: Thursday, August 9, 2012 - 1pm to 3pm
    Presenters:
    Alternate Location: Wu and Chen Auditorium (Levine 101)
    GRASP REU Site Oral PresentationsThursday, August 9, 2012 Wu and Chen Auditorium 1pm – 3pm   Welcome by Katherine J. Kuchenbecker and Max Mintz, GRASP REU Site Co-Directors 1:00 p.m.     Stephen Dodds Rising Senior in Electronics Engineering at the University of Nebraska at Omaha Advised by Dr. Vijay Kumar; mentored by Dr. Ed Steager and Denise Wong Developing Tracking Algorithms For Multiple Microrobots Abstract: Microrobotics is an exciting area for research because this new field became possible after recent technological advances in microstructure manufacturing and observation. Tracking the microrobots is an essential element in quantitative observation and control. Most of the existing tracking systems for microstrustures are limitted to tracking only one object, or robot, at a time. In this paper I introduce two methods for tracking multiple microrobots. Contour tracking was implemented first and the limitations were explored. Feature tracking was implemented second to make up for some of the limitations of the contour tracker. I show that feature tracking is accurate even when one robot overlaps another, when a robot drifts partially off the screen, or when clusters of bacteria collide with the robot. We show that contour tracking becomes less accurate when the previously stated occur. 1:15 p.m.     Lowell Fluke Rising Junior in Applied Mathematics at Harvard University Advised by Dr. Kostas Daniilidis and Dr. Vijay Kumar; mentored by Nicu Stiurca and Dr. Koushil Sreenath Autonomous Quadrotor Navigation With Visual Servoing Abstract: Among miniature aerial vehicles (MAV), quadrotor helicopters are becoming the most popular platform for research due to their versatility, and their ease of maintenance and design.  I consider the problem of autonomously flying a quadrotor using one on-board camera, with the goal of performing tasks such as flying through an opening and picking up objects.  Autonomous quadrotor flight using a monocular on-board camera is useful for exploration, rescue, and surveillance in unknown environments, and this paper explores the image based visual servoing (IBVS) controller scheme.  While previous approaches have attempted to estimate the position of the MAV before control is implemented or build a 3D model of the environment for planning, IBVS attempts to execute real-time control within the image space of the on-board camera.  While IBVS is faster and less computationally expensive, it has disadvantages such as the necessity of maintaining a view of the desired object.  I perform Matlab simulations to explore IBVS on a quadrotor and propose methods for testing this system. 1:30 p.m.     Rocky Foster Rising Senior in Mathematics at the University of Maryland, Baltimore County Advised by Dr. Alejandro Ribeiro; mentored by Jim Stephan Robust Fluctuating Computer Networks with Application to Robotic Systems Abstract: In this paper, maintaining link-to-link and end-to-end communications over a stochastic model of robots is discussed.  Using predetermined probabilities gathered by previous trials, simulations are developed to test methods and models of end-to-end communication systems.  This research is centered around a variation of TCP/UDP_IP communication systems.  The TCP/UDP_IP communication comes from the wireless communication of the robots and how to deal with packet loss and transfer.  The variation comes because probabilities must be re-calculated on a frequent basis as opposed to having a static system. 1:45 p.m.     Tre Glover Rising Sophomore in Computer Engineering at the University of Maryland, Baltimore County Advised by Dr. Camillo J. Taylor; mentored by David Isele Optimization Methods with Landmark Based Robot Navigation Abstract: This paper looks into the ability to navigate from one position to another based on common landmarks visible from both positions.  A scheme was used in order to generate the direction in which to travel between the two positions using the landmarks.  Also, a mapping method enables the navigation of longer distances by linking multiple local navigation processes.  In order to reduce the distance that is traveled and make the process more resilient to common errors, two other methods were implemented.  These methods all come together to create a single algorithm for landmark based navigation. 2:00 p.m.     Raven Hooper Rising Sophomore in Electrical and Computer Engineering at Temple University Advised by Dr. Katherine J. Kuchenbecker; mentored by Heather Culbertson Real-Time User Interface for the Haptic Camera Abstract: When you drag a tool across an object, you can feel many of the properties of its texture including roughness, hardness, bumpiness, and stickiness. The vibrations of the tool are responsible for a large portion of the feel of an object. Capturing the feel of real objects can be done with the use of the Haptic Camera, which records normal force, scanning speed, and tool vibrations as the user drags the tool across the textured surface. A magnetic tracking sensor, a force-torque sensor, and two 2-axis accelerometers held together with a pen-like covering is the Haptic Camera device. The tracker has 6DOF and uses x, y, z coordinates to find the position of the tool. The ATI Nano 17 force sensor measures the normal force of the Haptic Camera when it is pressed downward on an object. In order to get accurate data recordings, users need to record vibrations from the textures at several scanning speeds of up to 400 mm/s per second and several normal forces up to 4 Newtons. However, in the previous system, the user was not provided with any real-time indication of the speeds and forces they used during the recording. It was impossible to tell if the user had sufficiently varied scanning speed and normal force until after the recording sessions, which increased the amount of time and data required. I designed a feedback screen that measures scanning speed verse normal force. To provide a second visual indicator of force, an LED was attached to the ATI Nano 17 force sensor so as you press downward with the Haptic Camera the LED's light up accordingly.  2:15 p.m.     Liz Miller Rising Sophomore in Robotics Engineering at Worcester Polytechnic University (WPI) Advised by Dr. Vijay Kumar; mentored by Yash Mulgaonkar Optimizing Quadrotor Performance through Modeling and Scaling Abstract: ​With recent work in unmanned autonomous vehicles focusing mostly on computer vision and environment mapping, this project focuses specifically on improving quadrotor frames through material and scale modeling. Making quadrotors structurally optimal is important in order to increase performance and agility. Research primarily involved modeling the Hummingbird quadrotor frame and assigning variables to represent scale factors. Designers have a better idea on how to construct the best frame for responsive quadrotors using the relationships between moment of inertia, scale, and material selection. 2:30 p.m.     Chase Pair Rising Senior in Mechanical Engineering at the University of Texas, Arlington Advised by Dr. Vijay Kumar; mentored by Justin Thomas and Dr. Koushil Sreenath Obtaining Tension Measurements for Dynamic Quadrotor Load Sharing and Control Abstract: To enable dynamic control of the position and attitude of a shared load carried by two quadrotors, we focus on obtaining measurements of tension in the cables connecting each quadrotor to the load. Towards this, we show that the combined system of two quadrotors and cables is differentially flat and use this for designing dynamic trajectories. This will enable performing aggressive maneuvers while carrying the load which can be used to swing loads and quadrotors through obstacles. 2:45 p.m.     Julie Walker (in absentia) Rising Junior in Mechanical Engineering at Rice University Advised by Dr. Katherine J. Kuchenbecker; mentored by Will McMahan and Jennifer Hui Haptic Feedback for Remote Palpation Abstract: During open procedures, surgeons often palpate tissue with their fingers to determine its stiffness and explore for lumps of harder material, such as tumors. However, surgeons performing minimally invasive laparoscopic or robotic surgery use long thin tools, so they cannot directly touch the patient with their fingers. This project presents a fingertip haptic feedback device that seeks to replicate some of the touch sensations felt during palpation. The design focuses on replicating the feel of tissue stiffness (normal force), texture (vibrations), and the sensation of making and breaking contact with surfaces. These signals are measured by a sensorized palpation device, such as the SynTouch Biomimetic Tactile (BioTac) sensor. While holding the user’s finger in place, the feedback device uses a motor and a linkage to lift a surface up to the fingertip and apply the measured forces and vibrations.  Simultaneously, the force distribution detected by the sensor is displayed visually on a computer screen. In this way, the device allows one to see and feel the difference between hard and soft materials remotely. Attaching the touch sensor to a robotic surgery instrument and the feedback device to the control console would give surgeons the ability to palpate without directly touching the patient. Many thanks to the advisors, mentors, colleagues, staff, GRASP Lab, and larger Penn community for helping make the first year of the GRASP REU Site such a success.  We are especially indebted to Charity Payne and Robert Parajon for their excellent work in running the program this summer.  Congratulations to all eight 2012 Participants!
  • Talk: Fall 2012 GRASP Seminar - David Forsyth, University of Illinois at Urbana-Champaign, "Understanding Rooms: Geometry, Lighting and Material"
    Date: Friday, August 31, 2012 - 11am to 12pm
    Presenters: David Forsyth
    Alternate Location: Wu and Chen Auditorium (Levine 101)
    D.A. Forsyth, with Varsha Hedau, Derek Hoiem, and Kevin Karsch Rooms are important, because people live in rooms. Recent methods can now recover a reasonable, box-like approximate geometric model of a room from a single picture. This geometric model, while not necessarily exact, is surprisingly useful. For example, it can be used to enhance detection of furniture. I will show how relatively conventional detectors can be improved by taking this geometry into account. This box approximation can also be used to parse the room into occupied and free space. Free space is interesting because it has potential --- free space consists of volumes into which one could move, for example. However, free space is not empty. It is occupied by light traveling through the room. I will show how our box approximation allows us to extend relatively straightforward lightness and shading inference methods to produce relatively accurate estimates of illumination in space. These estimates can be used to light new objects and insert them into the room. I will show numerous compelling examples of objects inserted into legacy images of rooms while preserving an appearance of natural lighting. Finally, I show we can make estimates of the parameters (diffuse albedo, etc.) of the materials from which objects are made. This works because we know how much light is where in a room, and because we can see the light leaving objects.
  • Talk: Fall 2012 GRASP Seminar - Jean Gallier, University of Pennsylvania, "The Classification Theorem for Compact Surfaces"
    Date: Friday, September 14, 2012 - 11am to 12pm
    Presenters: Jean Gallier
    The classification theorem for compact surfaces is one of the great achievements of early 20th century mathematics. The statement of this theorem is quite intuitive but it took about sixty years until a rigorous proof was finally given by Brahana in 1921. Early versions of the classification theorem were given by Mobius in 1861, and by Jordan in 1866. More definite proofs were given later by von Dyck in 1888 and Dehn and Heegaard in 1907. This talk is about the history of the theorem and the techniques used to prove it. We will give a guided tour of the proof, pointing out which tools from algebraic topology are needed, and give an abbreviated  history of the "proof." A byproduct of the theorem yields  "global parametrizations," using fundamental domains, a recent topic of research.
  • Talk: Fall 2012 GRASP Seminar - Darius Burschka, Technische Universität München, "Embedded Visual Perception for Navigation and Manipulation"
    Date: Friday, September 21, 2012 - 11am to 12pm
    Presenters: Darius Burschka
    I will present the work of the Machine Vision and Perception Group at TUM in the field of navigation and object registration on a variety of systems including our flying and manipulation platforms. Sensing is essential for autonomy in robotic applications. Our focus is on how to provide sensing to low power systems that enables them to cope with the high dynamics of the underlying hardware. A disadvantage of using compact, low-power sensors is often their slower speed and lower accuracy making them unsuitable for direct capture and control of high dynamic motion. On the other hand, the inherent instability of some systems (e.g. helicopters or quadrotors), their limited on-board resources and payload, their multi-DoF design and the uncertain and dynamic environment they operate in, present unique challenges both in achieving robust low level control and in implementing higher level functions. We developed tracking algorithms (AGAST) and localization (Z_inf) techniques that can be used for navigation on embedded systems. I will show their application on OMAP3 processors (BeagleBoard.org system). Perception of the sensors can be boosted by adding external data in form of sensor data fusion or indexing to external databases. I will present an efficient 3D object recognition and pose estimation approach for grasping procedures in cluttered and occluded environments. In contrast to common appearance-based approaches, we rely solely on 3D geometry information. Our method is based on a robust geometric descriptor, a hashing technique and an efficient, localized RANSAC-like sampling strategy.
  • Talk: Fall 2012 GRASP Seminar - GRASP Faculty Research Projects, University of Pennsylvania, "GRASP Faculty Kick-Off Seminar"
    Date: Friday, September 28, 2012 - 11am to 12pm
    Presenters: GRASP Faculty
  • Talk: Fall 2012 GRASP Seminar - Petros Maragos, National Technical University of Athens, "Morphological and Variational Methods in Image Analysis and Vision"
    Date: Friday, October 5, 2012 - 11am to 12pm
    Presenters: Petros Maragos
    This talk presents an overview of some advances in two broad research directions in image analysis and computer vision that share in common the properties of being nonlinear and geometric. The first is based on morphological image operators and their lattice-theoretic generalizations which have a rich algebraic structure. The second approach uses nonlinear PDEs some of which are related to morphological operators and/or are derived from a variational formulation.  Both approaches and often their combination are useful for multiscale edge-preserving smoothing, feature detection, image simplification, structure+texture decomposition, segmentation, and shape analysis.  After a brief synopsis of morphological operators on images and graphs,  we shall continue with their PDE and variational formulation. First we focus on a class of multiscale connected operators with a combined local and global action where the PDE and the lattice approach can harmoniously work together, the first to provide continuous-scale isotropic-growth models with global constraints and the second to study discrete algorithms for numerical implementations. Then, we describe their usage for image simplification, structure+texture decomposition, and PDE-based image segmentation implemented by levelset curve evolution and driven both by watershed flooding and a texture oscillation energy. In an alternative scheme, this energy approach and a modulation image model help us develop an efficient unsupervised segmentation approach using region competition and weighted curve evolution based on probabilistic cue integration. If time permits, we shall also summarize some ongoing work in patch-based PDEs for tensor-based image diffusions using a variational framework.
  • Talk: Fall 2012 GRASP Seminar - John Enright, Kiva Systems, "Optimization and Coordinated Autonomy in Mobile Fulfillment Systems"
    Date: Friday, October 12, 2012 - 11am to 12pm
    Presenters: John Enright
    The task of coordinating hundreds of mobile robots in one of Kiva System's warehouses presents many challenging multiagent resource allocation problems. The resources include things like inventory, open orders, small shelving units, and the robots themselves. The types of resources can be classified by whether they are consumable, recycled, or scheduled. Further, the global optimization problem can be broken down into more manageable sub-problems, some of which map to (hard) versions of well known computational problems, but with a dynamic, temporal twist.
  • Talk: Fall 2012 GRASP Seminar - Ufuk Topcu, University of Pennsylvania, "Specification and Synthesis of Hierarchical Control Protocols and Some Applications in Autonomy and Robotics"
    Date: Friday, October 26, 2012 - 11am to 12pm
    Presenters: Ufuk Topcu
    The talk begins with the formulation of a control protocol synthesis problem and some background material. The solution discussed in the talk builds on temporal logics, two-player turn-based games, and optimization-based control. It yields to a partially automated method for constructing hierarchical control protocols. An autonomous navigation case study will demonstrate the results. Several other examples will be drawn from dexterous robotic manipulation and smart camera networks. The talk ends with an overview of some open issues biased to those from the speaker's own work.
  • Talk: IRCS / GRASP Seminar - Dieter Fox, University of Washington, "Grounding Natural Language in Robot Control and Perception Systems"
    Date: Friday, November 2, 2012 - 11am to 12pm
    Presenters: Dieter Fox
    Robots are becoming more and more capable at reasoning about people, objects, and activities in their environments. The ability to extract high-level semantic information from sensor data provides new opportunities for human robot interaction. One such opportunity is to explore interacting with robots via natural language. In this talk I will present our recent work toward enabling robots to interpret, or ground, natural language commands in robot control systems. We build on techniques developed by the semantic natural language processing community on learning combinatory categorial grammars (CCGs) that parse natural language input to logic-based semantic meaning. I will demonstrate results in two application domains: First, learning to follow natural language directions through indoor environments; and, second, learning to ground object attributes via weakly supervised training. Joint work with Luke Zettlemoyer, Cynthia Matuszek, Nicolas Fitzgerald, Yuyin Sun, and Liefeng Bo. Support provided by Intel ISTC-PC, NSF, and ARL, and ONR.
  • Talk: Fall 2012 GRASP Seminar - Ashutosh Saxena, Cornell University, "How Should a Robot Perceive the World?"
    Date: Friday, November 9, 2012 - 11am to 12pm
    Presenters: Ashutosh Saxena
    In order for a robot to perform tasks in the human environments, it first needs to figure out "what" to perceive. While for some robotic tasks, only geometry and semantic labels are good enough, many other robotic tasks require a robot to be more creative about what to perceive. For example, for a robot to arrange a disorganized room, it would need to perceive the human preferences about the usage of objects as well as the low-level manipulation strategies. In this talk, I will illustrate the issues surrounding "what to perceive" through a few examples. The key to figuring out "how" to perceive lies in being able to model the underlying "structure" in the problem. I propose that for reasoning about the human environments, it is the humans that are the true underlying structure in the problem. This is not only true for tasks that involve humans explicitly (such as human activity detection), but also true for tasks in which a human was never observed! In this talk, I will present learning algorithms that model such underlying structure in the problem. Finally, I will present several robotic applications ranging from single-image based aerial vehicle navigation to personal robots performing tasks of unloading items from a dishwasher, loading a fridge, arranging a disorganized room, and performing assistive tasks in response to human activities.
  • Talk: Special GRASP Seminar - Gabe Sibley, George Washington University, "Mobile Robot Perception for Long-term Autonomy"
    Date: Monday, November 12, 2012 - 2pm to 3pm
    Presenters: Gabe Sibley
    Alternate Location: Levine 512 (3330 Walnut Street)
    If mobile robots are to become ubiquitous, we must first solve fundamental problems in perception. Before a mobile robot system can act intelligently, it must be given -- or acquire -- a representation of the environment that is useful for planning and control. Perception comes before action, and the perception problem is one of the most difficult we face. An important goal in mobile robotics is the development of perception algorithms that allow for persistent, long-term autonomous operation in unknown situations (over weeks or more). In our effort to achieve long-term autonomy, we have had to solve problems of both metric and semantic estimation. In this talk I will describe two recent and interrelated advances in robot perception aimed at enabling long-term autonomy. The first is relative bundle adjustment (RBA). By using a purely relative formulation, RBA addresses the issue of scalability in estimating consistent world maps from vision sensors. In stark contrast to traditional SLAM, I will show that estimation in the relative framework is constant-time, and crucially, remains so even during loop-closure events. This is important because temporal and spatial scalability are obvious prerequisites for long-term autonomy. Building on RBA, I will then describe co-visibility based place recognition (CoVis). CoVis is a topo-metric representation of the world based on the RBA landmark co-visibility graph. I will show how this representation simplifies data association and improves the performance of appearance based place recognition. I will introduce the "dynamic bag-of-words" model, which is a novel form of query expansion based on finding cliques in the co-visibility graph. The proposed approach avoids the -- often arbitrary -- discretization of space from the robot's trajectory that is common to most image-based loop-closure algorithms. Instead, I will show that reasoning on sets of co-visible landmarks leads to a simple model that out-performs pose-based or view-based approaches, in terms of precision and recall. In summary, RBA and CoVis are effective representations and associated algorithms for metric and semantic perception, designed to meet the scalability requirements of long-term autonomous navigation.
  • Talk: Fall 2012 GRASP Seminar - Visa Koivunen, Aalto University, Finland, "Optimal Array Processing in The Face of Nonidealities (with wireless localization examples)"
    Date: Friday, November 16, 2012 - 11am to 12pm
    Presenters: Visa Koivunen
    In this talk we describe techniques that allow the practitioner to apply high performance array processing algorithms using real-world sensor arrays with nonidealities. Arbitrary array geometries and processing data  in azimuth, elevation and polarimetric domains are considered. We acquire a realistic array steering vector model by taking into account array nonidealities such as mutual coupling, mounting platform reflections, cross-polarization effects, errors in  element positions as well as individual directional beampatterns. This facilitates achieving optimal or close-to-optimal performance and retaining high-resolution capability despite the nonidealities. Moreover, tighter performance bounds may be established for parameter estimation. We describe how the various  approaches can be applied in practice in the context of high-resolution  direction finding as well as beamforming so that problems related to beamsteering, SOI and interference cancellation are mitigated.
  • Talk: Fall 2012 GRASP Seminar - Jason Corso, University of Buffalo, "Advances in Segmentation for Video Understanding"
    Date: Friday, November 30, 2012 - 11am to 12pm
    Presenters: Jason Corso
    The use of video segmentation as an early processing step in video understanding lags behind the use of image segmentation for image understanding, despite many available video segmentation methods. The reasons for this are likely due to a general lack of critical analysis to help us understand which methods work well in which scenarios, and the simple fact that videos are an order of magnitude bigger than images. In this talk, I will cover recent advances in my group that address both of these reasons. First, I will discuss LIBSVX, a suite of five supervoxel methods coupled with a set of spatiotemporal metrics to evaluate the segmentation methods. Our evaluation summarily arrives at the conclusion that hierarchical segmentation methods, which reevaluate similarity at multiple scales within the hierarchy, perform best overall. Despite the hierarchical methods performing best, the state of the art hierarchical algorithms have two significant limitations: they require the full video to be loaded into memory, limiting the videos they can process, and the hierarchical output they produce often overwhelms the user with too much data; it is not always clear which level of the hierarchy should be used. The second part of the talk will discuss how we systematically address these two limitations. We have proposed an approximation framework for streaming hierarchical video segmentation motivated by data stream algorithms: each video frame is processed only once and does not change the segmentation of previous frames. We also propose a new criterion for flattening the hierarchy based on the notion of uniform motion entropy; select segments throughout the hierarchy so as to balance the amount of motion entropy within the selected segments. Time permitting, I will present an example of how video understanding can benefit from using the segmentation: for the video label propagation problem, our supervoxel-based propagation method is significantly more capable than the best of the state of the art pixel-based methods.
  • Talk: Special GRASP Seminar - Elmar Mair, German Aerospace Center (DLR), "Efficient and Robust Navigation Based on Inertial and Visual Sensing"
    Date: Friday, November 30, 2012 - 2pm to 3pm
    Presenters: Elmar Mair
    Alternate Location: Levine 307 (3330 Walnut Street)
    Reliable motion estimation on resource-limited platforms is a crucial task for many applications. While insects solve this problem in an exemplary manner, mobile robots still require a bulky computation and sensor equipment to provide sufficient robustness. In this talk, I will motivate the use of an inertial-visual system as a minimal sensor concept which still allows an efficient and robust navigation. I will focus on image processing, especially efficient feature tracking and motion estimation algorithms, resulting in an algorithm which tracks several hundreds of features in a few milliseconds on low-power processing units. Such high frame-rates are of great interest if high dynamic mobile robots, like multicopters, have to be controlled. A perturbation analysis of motion estimation algorithms gives insights how the resulting accuracy depends on the aperture angle, the tracking accuracy, and the number of features. Furthermore, an algorithm is presented which accurately aligns sensors like an IMU and a camera in time and in space. Finally, I will discuss an insect-inspired navigation concept, which enables long-distance navigation even on memory limited systems. Applications will be presented like the DLR 3D-Modeler and the DLR Multicopters, where these methods are applied to.
  • Talk: GRASP Special Seminar - Avi Ziskind, Columbia University, "Neurons in Cat V1 cluster by degree of tuning but not by absolute spatial phase or temporal response phase"
    Date: Friday, December 14, 2012 - 11am to 12pm
    Presenters: Avi Ziskind
    Alternate Location: Levine 512 (3330 Walnut Street)
    Nearby neurons in cat primary visual cortex (V1) have similar preferred orientation, direction, and spatial frequency. How diverse is their degree of tuning for these properties? Are they also clustered in their tuning for the spatial phase of a flashed grating or the temporal phase of a drifting grating? To address these questions, we used tetrode recordings to simultaneously isolate multiple cells at single recording sites and record their responses to gratings of multiple orientations, spatial frequencies, and spatial/temporal phases. We found that orientation tuning width, spatial frequency tuning width and direction selectivity index all showed significant clustering. Tuning for the spatial phase of a flashed grating stimulus (“absolute spatial phase”) and temporal phase of a drifting grating stimulus (“temporal phase”) however, showed no clustering.