The rapid growth of digital libraries (DLs) worldwide poses many new challenges for document image analysis (DIA) research and development. DLs promise to offer more people access to larger document collections, and at far greater speed, than physical libraries can.  But DLs also tend, for many reasons, to serve poorly, or even to omit entirely, many types of non-digital human--legible media such as originally printed and handwritten documents.   These documents, in their original physical (undigitized) form, are readily - if not always quickly - legible, searchable, and browseable, whereas in the form of images accessed through DLs they often lose many of their original advantages while of course lacking many advantages of symbolically encoded information.  This talk explores these issues and illustrates them with brief case studies arising from his experience as a DIA researcher in collaboration with several DL projects in the US.  Difficult open DIA technical problems in DL applications are identified in, e.g., the contrasting advantages of paper and digital displays, during image capture, early processing, recognition, analysis, presentation, and retrieval - and in personal and interactive applications.

These support the conclusion that the international image understanding R&D community is urgently needed (because uniquely qualified) to provide new technology to help rescue from neglect - even, in many cases, eventual oblivion - the world's vast culturally irreplaceable legacy paper document collections.

Peter J. Burt
Sr. Director,

Vision Technologies, 

Sarnoff Corporation

Abstract:

The science and art of computer vision has matured over the past several years to the point that vision systems can guide robotic vehicles in real time through complex natural outdoor environments. Capabilities have been developed to analyze scene structure from on-board moving video cameras, maintain precision vehicle location and pose, and, not least, provide high performance, low latency vision processing within the vehicle control loop.  I will describe contributions made at Sarnoff to these achievements.  Current capabilities mimic a human’s ability to navigate using his own eyes for visual guidance.  I will also describe our work leading to future capabilities in which robots can share the “eyes” of other robots to achieve “distributed vision” and iconic memory of the world they move through.

Ken Nakayama

Vision Sciences Laboratory
Department of Psychology
Harvard University, USA.

Title:

Perception and recognition of faces – human capacities of possible relevance for artificial systems.

Abstract:

Faces are one of the most frequently looked at objects in our environment.  Our eyes scan faces for myriad reasons, to assess identity, health, moods, intentions, imminent actions, to name a few.   As such, it should not be surprising that face processing is a well developed function in the human brain with the likelihood of dedicated hardware.  We take as a working hypothesis that the processing of faces is unique, that it is qualitatively different from the processing of other objects.  

We have examined this idea using a variety of methods, including fMRI, magnetoencephalography as well testing visual capacities using psychological and psychophysical methods.   We have concentrated on two different groups, normal observers and those suffering from prosopagnosia.  This latter group is of great interest because of their selective deficiency in face recognition.  We document this by showing that these individuals are very poor at remembering faces but that they can recognize other objects normally.  A fraction of these individuals show abnormal neural responses to faces. 

 We also show that in normal observers, the human face recognition system is selectively sensitive to small facial features, approximately on a scale of 0.5 centimeters on the face, that the processing of upright faces is holistic whereas upside down faces and all other objects are encoded in a piecemeal fashion.  These findings reinforce the idea that the recognition of faces is qualitatively different from that of objects.

We also show that face recognition is not easily degraded by affine image transformations, suggesting that for purposes of recognition, faces can be regarded as flat two dimensional objects.

Brian Curless

Department of Computer Science & Engineering

University of Washington

Washington 98195-2350

USA

Title:

The Space of Human Shapes

Abstract:

The human form has been the dominant subject of painters and sculptors for millennia.  It is the shape we are most attuned to, the one most important to recognize and to build our environments around.  Modeling this shape and the way it moves has been a grand challenge in computer science.  This task is one of the main components of the graphics Turing test: can a machine fool an observer into thinking the synthetic projection of a moving, talking person is real.

In this talk, I present strides taken in the direction of modeling realistic, time-varying human shape.  These strides are formed around the premise that realism depends on samples from reality.  The capacity to sample the shape of the real world has grown dramatically in recent years with the advent of fast, reliable imaging sensors and controllable illumination sources.  Using existing shape scanners, I will describe how my collaborators and I have developed tools for modeling how a human shape varies over time and how body shape varies across a population.  I will also present recent work on modeling the most challenging part of a human: the face.  For this task, we have developed a completely new shape scanner that enables us to capture a moving face at high spatio-temporal resolution.  To illustrate the success of these shape capture and modeling methods, I will give live demonstrations of interactive human body modeling and three-dimensional video reconstructions of a human face.  Finally, I will discuss open problems on the frontiers of digital human modeling.

Partha Pratim Das

Interra Systems

India