Keynotes

Large Online Image Collections as a Source
of Visual Realism

Alexei Efros, Carnegie Mellon University, USA
Tuesday, May 4, 10:00–11:00 in Concert Hall

Abstract  —  Despite recent advances, visual realism in complex, real-world scenes remains largely elusive for traditional computer graphics. The main culprit appears to be the sheer complexity of our visual world -- a realistic scene requires much more detailed geometry and lighting than could possibly be provided by hand. Of course, if one happens to have access to the physical site to be modelled (e.g. a monument or a movie set), then image-based modeling and rendering techniques can be employed in situ, producing excellent results. However, it is often not possible to travel "on location" in order to capture the needed geometric and photometric data in person. Fortunately, over the last few years, the Internet has developed into a gargantuan depository of visual data (photos, videos, webcams, etc) captured by people all over the globe. While the recent work on Photo Tourism demonstrates the power of using online photo collections for modeling a specific site, most of the online imagery is, in fact, quite "generic". A pressing research question is how this generic visual data could be useful in graphics as a way of "crowd-sourcing" visual realism?

In this talk, I will give an overview of some of the recent work (both from our lab and elsewhere) on using large online image collections to transfer visual appearance as a way of synthesizing novel visual content. I will also discuss the dangers and pitfalls inherent in using very large datasets, including various sources of bias.

Biography  —  Alexei (Alyosha) Efros is an assistant professor at the Robotics Institute and the Computer Science Department at Carnegie Mellon University. His research is in the area of computer vision and computer graphics, especially at the intersection of the two. He is particularly interested in using data-driven techniques to tackle problems which are very hard to model parametrically but where large quantities of data are readily available. Alyosha received his PhD in 2003 from UC Berkeley and spent the following year as a post-doctoral fellow in Oxford, England. Alyosha is a recipient of CVPR Best Paper Award (2006), the NSF CAREER award (2006), the Sloan Fellowship (2008), the Guggenheim Fellowship (2008), and the Okawa Grant (2008).

Super-linear Slowdown with Multicores

Erik Hagersten, University of Uppsala, Sweden
Thursday, May 6, 16:00–17:00 in Concert Hall

Abstract  —  The recent trend with multicore architectures for general-purpose processing brings the promise of great performance at low cost. All you have to do is to parallelize your single application or run several independent applications simultaneously in a throughput manner.

However, multicores also introduce new bottlenecks that could result in slower-than-sequential execution even if your application is embarrassingly parallel. Tuning for multicore requires insights into complex resource sharing (e.g., caches and bandwidth) and thread interaction (e.g., synchronization, coherence and false sharing). It is not obvious that the computing capabilities of a chip will be determining its performance in the future. Who cares how many FLOPS, MIPS, cores, or threads a chip has if its performance is limited by other factors?

This talk will discuss the new challenges associated with multicore performance today and tomorrow. A couple of very simple optimizations will be shown to have dramatic impact on performance and scalability. Multicore may not bring any free lunch, but there is often plenty of low-hanging fruit.

Biography  —  Professor Erik Hagersten holds a professorial chair in computer architecture at Uppsala University Sweden since 1999. He is also the CTO of Acumem AB, developing new technology for multicore optimization.

He gained his initial interest in parallel architecture at the Dataflow project at MIT in the early 80s and applied parallel processing and programming at Ericsson in the late 80s. He coined the term "Cache-Only Memory Architecture" (and its brain-dead acronym, COMA) while managing the architecture research group at the Swedish Institute of Computer Science in the early 90s. The Simics simulator is another result from that group.

Hagersten was the chief architect for Sun Microsystem's high-end server engineering division from 1993 to 1999. He was the architect of the Sun WildFire, the Sunfire Link, the Sun Enterprise 15k/25k and UltraSPARC III and IV scalable coherence technology.

More than a Quick User Study:
Perception in Computer Graphics

Erik Reinhard, University of Bristol, UK
Friday, May 7, 11:00–12:00 in Concert Hall

Abstract  —  Human visual perception is playing an increasingly important role in computer graphics. This often takes the form of a rudimentary perceptual study to validate the utility of a given algorithm. Nonetheless, this field has several further implications for computer graphics. Aside from giving tools to measure visual quality, there exist computational models of human vision which can be directly leveraged, particularly in tone reproduction and colour appearance modeling. We have also gained insights in the link between natural images and human vision, which can be useful in several applications such as deblurring images and colour transfer.

Further, human visual perception allows us to solve computational problems that would otherwise be very difficult to tackle. For instance, many high-level image edits require an analysis of the image, recovering a basic understanding of the shape of objects. On the basis of a single image, this is an under-constrained problem. To make headway in solving such problems, human visual perception has emerged as a powerful tool. By requiring perceptual plausibility rather than physical accuracy, practical solutions to such image edits are possible. This has led to several interesting applications, including image-based material editing.

In summary, the utility of visual perception goes much further than incorporating the obligatory perceptual study in many graphics papers. This point will be illustrated by means of practical examples.

Biography  —  Erik Reinhard received his Ph.D. in Computer Science from the University of Bristol in 2000, having worked on his Ph.D. for three years at Delft University of Technology, prior to a further three years in Bristol. Following a post-doctoral position at the University of Utah (2000-2002) and an assistant professorship at the University of Central Florida (2002-2005), he returned to Bristol as a lecturer in January 2006 to become senior lecturer in 2007. Erik founded the prestigious ACM Transactions on Applied Perception, and was Editor-in-Chief from its inception in 2003, until early 2009. He is lead author of two books: 'High Dynamic Range Imaging: Acquisition, Display, and Image-Based Lighting' and 'Color Imaging: Fundamentals and Applications'. His interests are in the application of knowledge from perception and neuroscience to help solve problems in graphics and related fields.