Winter 2007 Talk Descriptions

Phenotype and functional discovery using a lentiviral RNAi library and automated imaging

Jason Moffat

Abstract: Systematic knockdown of gene function in human cells has been made possible by the discovery of RNA interference. To enable arrayed or pooled loss-of-function screens in a wide range of mammalian cell types, I have helped develop genome-wide lentiviral short hairpin (shRNA) libraries targeting human and mouse genes as part of a large academic/industry consortium. To test the utility of the library for arrayed screens, we used a pilot set of >5000 lentiviruses and designed a screen based on high-content imaging to identify genes required for mitotic progression in human cancer cells. The screen identified several known regulators and many candidate regulators of mitotic progression and proliferation. By combining systematic lentiviral mediated RNAi and high content screening we are building libraries of images representing different knockdown conditions and can query phenotypic signatures using novel classification tools in order to assemble rudimentary genetic networks. Harnessing the power of functional genomics in mammalian cell culture will accelerate our understanding of signaling pathways and help dilineate novel therapeutic targets.

Problems in Autonomous Devices

Steve Cousins

Abstract: Autonomous devices are just appearing on the horizon. Roomba and similar automated vacuum cleaners are a moderately successful commercial novelty. In 2005 five teams built autonomous vehicles that drove 132 miles across the desert. Robotics is moving from closely-controlled environments to more open interaction with people. Making this happen will require technologies that were previously only shown in labs to become "real", and will uncover challenging new problems to solve. Willow Garage is a new research lab that aims to uncover and solve these problems by building autonomous devices.

Steve Cousins is President and CEO of Willow Garage, Inc. Prior to joining Willow, he was a senior manager at IBM Research, and a member of the senior staff at Xerox PARC. He holds a PhD from Stanford University.

Satisfying Computer Vision Task Requirements in Video Camera Networks

Stan Sclaroff

Abstract: In the first part of the talk, I will describe a solution to the off-line problem of determining camera locations in a given layout such that the needs of specific computer vision tasks (resolution, coverage, etc.) are satisfied, while minimizing cost constraints. Researchers in Computational Geometry have proposed elegant solutions for some sensor location problem classes, but employ unrealistic assumptions about the cameras' capabilities that make the solutions unsuitable for many real world computer vision applications. We first define a general camera placement problem with assumptions that are consistent with the capabilities of real cameras then formulate a solution via binary optimization over a discrete problem space. We demonstrate the effectiveness of these ideas in experiments that test several real floorplan configurations and real camera types.

In the second part of the talk, I will describe a solution to the on-line problem. Given a camera layout containing active, e.g., pan-tilt-zoom, cameras and people moving around, our goal in the on-line phase is to predict a subset of cameras, respective camera parameter settings, and time windows that will most likely lead to success of particular vision tasks when a camera observes an event of interest. We propose an adaptive probabilistic model which accrues temporal camera correlations over time as the cameras report observed events. No extrinsic, intrinsic or color calibration of cameras is required. The camera and parameter prediction is achieved efficiently by a Sequential Monte Carlo sampling across space and time. We demonstrate the performance of the model in both simulated and real environment experiments using several active cameras.

This work is done in collaboration with Ugur Murat Erdem.

Learning Structured Appearance Models from Captioned Images of Cluttered Scenes

Mike Jamieson

Abstract: We address the problem of learning both the semantics (names) and the visual features (SIFT collections) of objects appearing in a training set of unstructured, captioned images of cluttered scenes. Prior work in applying machine translation models to learn the associations between image features and caption nouns has assumed a one-to-one correspondence between features and nouns. However, each training image may contain thousands of SIFT features belonging to multiple objects. Our challenge is two-fold: 1)~grouping the SIFT features into meaningful configurations, and 2)~learning the object names associated with those configurations. Since better feature configurations tend to have stronger associations with object names, we offer an integrated solution that uses the caption words to drive the feature grouping process. The result is a more general model acquisition framework that does not assume words correspond to individual features and does not require training images with isolated objects or unambiguous labels.

Learning Hierarchical Shape Models from Examples

Alex Levinshtein

Abstract: We present an algorithm for automatically constructing a decompositional shape model from examples. Unlike current approaches to structural model acquisition, in which one-to-one correspondences among appearance-based features are used to construct an exemplar-based model, we search for many-to-many correspondences among qualitative shape features (multi-scale ridges and blobs) to construct a generic shape model. Since such features are highly ambiguous, their structural context must be exploited in computing correspondences, which are often many-to-many. The result is a Marr-like abstraction hierarchy, in which a shape feature at a coarser scale can be decomposed into a collection of attached shape features at a finer scale. We systematically evaluate all components of our algorithm, and demonstrate it on the task of recovering a decompositional model of a human torso from example images containing different subjects with dissimilar local appearance.

Joint work with Cristian Sminchisescu and Sven Dickinson.

Virtual Environments to Assist Surgical Planning and Guidance

Terry Peters

Abstract: Surgical procedures that aim to treat diseases, often have the unfortunate side-effect of causing the patient significant trauma while accessing the target site. Indeed, in some cases the trauma inflicted on the patient during access to the target greatly exceeds that caused by performing the therapy. We have developed techniques for performing minimally-invasive surgery on the brain and heart, that rely on pre-operative images, combined with data acquired during the procedure. If we localise the region in the deep brain that causes Parkinson’s tremor, and represent it in a virtual model of the brain, instead of performing a craniotomy, we can accurately approach the target region with an ablation device or stimulator through a small burr-hole, guided by pre-operative images onto which prior knowledge of population electrophysiology has been mapped. In the heart, many intracardiac interventions are currently performed after the chest has been opened, the patient placed on cardiopulmonary bypass, and the heart arrested. We have developed approaches for operating on multiple targets inside the beating heart in a minimally-invasive fashion. In these procedures, the targets are accessed using a virtual reality (VR)-assisted image guidance platform that combines real-time ultrasound (US) with a virtual model of the surgical instruments, in the context of the 3D cardiac anatomy.

A Flux Invariant for Biological Shape

Kaleem Siddiqi

Abstract: In the late 60's Blum developed the notion of axis-morphologies for describing 2D and 3D forms. He proposed an interpretation of the local reflective symmetries of an object as as a "medial graph" and suggested that the implied part structure could could be used for object categorization and recognition. In this talk I will discuss a type of "flux" integral performed on the gradient vector field of the Euclidean distance function to the bounding curve (or surface) of an object. Remarkably, the limiting behavior of this integral as the enclosed area (or volume) shrinks to zero reveals a scalar invariant which both determines the Blum skeleton as well as the geometry of the object that it describes. I will also discuss our work on algorithms for computing medial loci using these ideas as well as the connection to some of the psychophysical literature on medial loci. Much more is covered in my upcoming book with Steve Pizer: "Medial Representations: Mathematics, Algorithms and Applications" (Springer, 2007, in press).

Hierarchically Learned Representations of Object Categories: From Pixels towards Semantic Parts

Ales Leonardis

Abstract: The question how to represent visual information in an artificial cognitive system to enable fast and reliable execution of various cognitive tasks has been discussed throughout the history of computer vision. The theories have converged towards hierarchical architectures of parts composed of parts, (the so-called compositional systems), starting with simple, frequent features that are gradually combined into more and more complex entities. However, the automatic design of parts in hierarchical layers has been hindered by a theoretically enormous number of possible compositions. In this talk, I will describe a novel approach that overcomes the exponential complexity of unsupervised learning by exploiting the favorable statistics of natural images in a sequential, hierarchical manner. The parts recovered in the individual layers of the hierarchy vary from simple to more complex ones and enable a fast indexing (bottom-up) and matching (top-down) scheme that can be efficiently used for a variety of cognitive tasks. I will show the results of the proposed approach obtained on different data sets, yielding important insights for designing compositional systems.

Continuous-state Graphical Models for Object Localization, Pose Estimation and Tracking

Leonid Sigal

Abstract: Reasoning about pose and motion of objects, based on images or video, is an important task for many machine vision applications. Estimating the pose of articulated objects such as people and animals is particularly challenging due to the complexity of the possible poses yet has applications in computer vision, medicine, biology, animation, and entertainment. Realistic natural scenes, object motion, noise in the image observations, incomplete evidence that arises from occlusions, and high dimensionality of the pose itself are all challenges that need to be addressed.

In this talk a class of approaches that model objects using hierarchical continuous-state graphical models will be presented. These approaches can be used to effectively model complex objects by allowing tractable and robust inference algorithms that are able to infer pose of these objects in the presence of realistic appearance variations and articulations. In these models that can be used to model both rigid and articulated object structures, nodes correspond to parts of objects and edges represent the constraints between parts encoded as statistical distributions. For articulated objects, these constraints can model spatial, temporal and occlusion relationships between parts. Localization, pose estimation, and tracking can then be formulated as inference in these graphical models. This formulation has a number of advantages over more traditional methods and can be used to solve the challenging problem of inferring the 3D pose of the person from single monocular image.

Computer Vision Problems in Abdominal Imaging

Masoom A. Haider

Abstract: The scope of clinical applications of diagnostic imaging has increased dramatically in te last decade. There has been a multifold increase in the variety and quantity of imaging data obtained for a single exam which has resulted in improved diagnostic ability. There has also been a shift to a pure digital environment from image acquisition to interpretation. There is a growing need for efficient methods of image analysis for segmentation, co-registration, 3D visualization, minimally invasive therapy delivery and computer aided diagnosis. Specific examples that will be discussed include the diagnosis, treatment and follow-up of renal masses; therapeutic response assessment of metastatic disease; and localization and image guided therapy of prostate cancer.

Deformable Model Segmentation by Posterior Optimization Based on Geometric Statistics and Regional Intensity Pattern Statistics

Stephen Pizer

Slides here

Abstract:For many years we have been developing a variety of methods that together would allow segmentation of 3D objects from medical images in a way reflecting knowledge of both the population of anatomic geometries sought and the population of images that were consistent with that geometry. The methods use the m-rep as the object representation and regional intensity quantile functions as the representation of image information in regions relative to the m-rep. Using manually segmented images to which m-reps have been fit and which contain information to allow alignment, our methods use principal geodesic analysis to estimate prior probability density, on the anatomic geometry, and they use principal component analysis to estimate a likelihood density, on the regional intensity quantile functions. They then segment automatically via posterior optimization over principal geodesic coefficients, after initialization via landmarks. Each component of this methodology will be briefly reviewed and compared to previous methods. Kidneys from multi-patient populations and pelvic organs from multi-day populations from a single patient were segmented by training a prior and a likelihood density by the methods indicated. Comparison to human segmentations provides good or better agreement with humans than the humans agree with each other.