Date & Time: 
Monday 29 September 2014

Methods for automated acquisition are offering new and exciting opportunities for large-scale acquisition of data samples for the sciences. Such large-scale data collections can be too massive to be realistically annotated by human operators, which creates a need for automated annotation. Conveniently, methods for automated semantic interpretation of images and audio samples are becoming increasingly powerful. This is evident in the ever-increasing applications of speech and image recognition for consumer products and businesses. However, as opposed to mainstream applications of automated identification, scientific applications often require statistical guarantees on the derived data-products.

In this seminar, Oscar will present two methods towards this goal. First, the problem of estimating a population mean using random sampling is considered. A hybrid sampling design is proposed that combines cheap sample collection, cheap-but-noisy automated annotation and accurate-but-costly expert annotation in a way that minimizes the total sampling cost. Second, methods for semi-automated and interactive annotation of coral reef survey images are presented. As demonstrated on survey images from four Pacific locations, these methods can significantly reduce the annotation time while maintaining the high quality of expert annotations.

About the presenter

Oscar Beijbom is a Ph.D Candidate at the Computer Vision group at the UCSD computer science department. His research focuses on the automated annotation and analysis of scientific data. Specifically, it is focused on the automated annotation of coral reef survey images, within the scope of a US National Science Foundation project: Computer Vision Coral Ecology. Recently, Oscar has been involved in the Catlin Seaview Survey project, where he works on methods for automated annotation of the collected imagery.

Before joining UCSD, Oscar was lead developer at Hövding, a Swedish startup company. As their first employee and engineer, Oscar created the algorithmic framework and hardware design for Hövding's renowned invisible bicycle helmet.

Oscar received his joint B.S / M.S degree in Engineering Physics from Lund University in 2007. His masters thesis work at Cellavision on Single Image Focus Level Assessment Using Support Vector Machines received several awards and generated an international patent.

GCI Seminar Room 275

Connect with us