Sunday, May 11, 2008

Moments from WIAMIS 2008

Keynote of H. Bischof

Recently, there has been considerable amount of research in methods for person detecti on. This talk will focus on methods for person detecti on in a surveillance setting (known environment). We will demonstrate that in this setti ng one can build robust and highly reliable person detectors by using on-line learning methods. In parti cular, I will fi rst discuss, conservative learning“ which is able to learn a person detector without any hand labelling effort. As a second example I will discuss a recently developed grid based person detector. The basic idea is to considerably simplify the detecti on problem by considering individual image locations separately. Therefore, we can use simple adapti ve classifi ers which are trained on-line. Due to the reduced complexity we can use a simple update strategy that requires only a few positive samples and is stable by design. This is an essential property for real world applications which require operati on for 24 hours a day, 7 days a week. During the talk we will illustrate our results on video sequences and standard benchmark databases.

Keynote of John R. Smith: Unleashing Video Search

Video is rapidly becoming a regular part of our digital lives. However, its tremendous growth is increasing users’ expectati ons that video will be as easy to search as text. Unfortunately, users are sti ll fi nding it diffi cult to fi nd relevant content. And today’s soluti ons are not keeping pace on problems ranging from video search to content classifi cati on to automati c fi ltering. In this talk we describe recent techniques that leverage the computer’s ability to eff ecti vely analyze visual features of video and apply stati sti cal machine learning techniques to classify video scenes automati cally. We examine related eff orts on the modeling of large video semantic spaces and review public evaluati ons such as TRECVID, which are greatly facilitati ng research and development on video retrieval. We discuss the role of MPEG-7 as a way to store metadata generated for video in a fully standards-based searchable representati on. Overall, we show how these approaches together go a long way to truly unleash video search. (Book af Abstracts)

No comments: