Fall 2009 Talk Descriptions

Fragmented Lenses and High ISO for Efficient Photography

Sam Hasinoff

Abstract: I'll describe two projects that address basic technical challenges in photography: (1) minimizing defocus blur, and (2) capturing high dynamic range. In both cases we characterize fundamental limits, and propose new methods which improve efficiency over the state-of-the-art.

First, I'll describe our new lens design, the "lattice-focal" lens, that can capture in-focus images over a greater range of depths than previous approaches. The design follows from our analysis of lens defocus over the 4D space of light rays. As we show, the only usable energy lies on a 3D subset of this space in the Fourier domain. We establish an upper bound on performance (ie. over any possible lens design), and show that the lattice-focal lens is closer to this bound than any previous design.

Second, I'll show how existing cameras can be used more efficiently, to capture high dynamic range scenes. For a given scene and camera, our analysis lets us compute the optimal sequence of photos to capture, maximizing worst-case SNR. This provides significant gains over standard exposure bracketing, typically 10 dB better or 3 times faster, when capture time is limited. As I'll explain, most our gains come from using high (but varying) ISO settings -- counterintuitively, "turning up the amplifier" can help reduce noise.

Bio: Sam Hasinoff received the BSc degree in computer science from the University of British Columbia in 2000, and the MSc and PhD degrees in computer science from the University of Toronto in 2002 and 2008, respectively. He is currently an NSERC Postdoctoral Fellow at the Massachusetts Institute of Technology. In 2006, he received an honorable mention for the Longuet-Higgins Best Paper Award at the European Conference on Computer Vision. He is the recipient of the Alain Fournier Award for the top Canadian dissertation in computer graphics in 2008.

Symmetry

Yanxi Liu

Abstract: Symmetry is an essential mathematical concept, as well as a ubiquitous, observable phenomenon in nature, science and art. Either by evolution or by design, symmetry implies an efficiency coding that makes it universally appealing, especially so to computational science. Recognition and categorization of symmetry and regularity is the first step towards capturing the essential skeleton of a real world problem, while at the same time minimizing computational redundancy. However, symmetry group detection from real world data turns out to be a challenging problem that has been puzzling computer vision, computer graphics and psychology researchers for decades. We explore a formal and computational characterization of real world regularity using a hierarchical model of symmetry groups as a theoretical basis, embedded in a well-defined Bayesian framework. Such a formalization simultaneously facilitates (1) a robust and comprehensive algorithmic treatment of the whole regularity spectrum, from regular (perfect symmetry), near-regular (approximate symmetry), to various types of irregularities; (2) an effective detection scheme for real world symmetries and symmetry groups; and (3) a set of computational bases for measuring and discriminating quantified regularities on diverse data sets. Besides some theoretical background on crystallographic groups in particular, I shall illustrate various applications of computational symmetry in texture synthesis, analysis, tracking, and manipulation; human gait and activity recognition; symmetry-based dance analysis; grid-cell clustering; automatic geo-tagging; and image ‘de-fencing’.

BIO: Yanxi Liu received her B.S. degree in physics/electrical engineering in Beijing and her Ph.D. degree in computer science for group theory applications in robotics from University of Massachusetts (Amherst). Her postdoctoral training was at LIFIA/IMAG (France). She also spent one year at DIMACS (NSF center for Discrete Mathematics and Theoretical Computer Science) under an NSF research-education fellowship award. Dr. Liu was with the research faculty in the Robotics Institute (RI) of Carnegie Mellon University before she joined the Computer Science Engineering and Electrical Engineering departments of Penn State University in Fall of 2006 as a tenured faculty and the co-director of the lab for perception, action and cognition (LPAC). Dr. Liu's research interests span a wide range of applications including computer vision, computer graphics, robotics, human perception and computer aided diagnosis in medicine, with two main themes: computational symmetry/regularity and discriminative subspace learning. Dr. Liu chaired the First International Workshop on Computer Vision for Biomedical Image Applications (CVBIA) in conjunction with ICCV 2005. Dr. Liu served as an area chair or organizing committee member for CVPR08/MICCAI08/CVPR09, and has served as a multi-year chartered study section member for the US National Institute of Health (NIH). Dr. Liu is a senior member of IEEE and the IEEE Computer Society.

Communicating 3D Shape using Lines

Doug DeCarlo

Abstract: The last few years have brought a more concrete understanding of the mathematical relationship between strokes in drawings and lines on 3D shapes. However, fundamental questions remain unanswered about how our perceptual system resolves these lines as giving evidence about shape. These questions need to be addressed to assemble mathematically defined lines into clear and compelling drawings. I'll discuss my ongoing research on suggestive contours and highlight lines, as well as perceptual studies that move us towards answering the question of what shape people see when they look at a line drawing.

Building Hierarchical Scene Representations from and for Human-Robot Interaction

Sven Wachsmuth

Abstract: The ultimate goal of human-robot interaction is to enable the robot to seamlessly communicate with a human about natural everyday environments. While most research in this area is concentrating on the communicative cues itself, it is frequently underestimated that the success of communication heavily relies on the compatibility of the representations behind it. If a speaker refers to an object or scene structure that the robot does not perceive or perceives differently, the robot cannot react, appropriately. In my talk, I will discuss different approaches how relevant scene structure (like functional room areas, tables, shelfs, doors, etc.) can be learned from coarse shape representations and human-robot interaction. The techniques are based on the processing of depth data and include holistic representations, the analysis of scene changes over time, verbal descriptions, and mixed-initiative dialog.