Beyond Classification -- Large-Scale Gaussian Process Inference and Uncertainty Prediction

Wednesday, December 12, 2012

DAAD scholar Erik Rodner, along with colleagues Alexander Freytag, Paul Bodesheim, and Joachim Denzler of the University of Jena, won the best paper honorable mention award at the Asian Conference on Computer Vision, held November 5-9 in Daejeon, South Korea. Their work tries to go one step further in lifelong visual learning with minimal supervision, an important topic in computer vision and robotics. Detecting new object categories in images and videos requires measuring classification uncertainty. The paper proposes methods of computing and approximating uncertainties in a Bayesian setting, which are in general intractable with large-scale data. Erik and his colleagues were able to show that the computation time can be reduced from several hours to milliseconds. Furthermore, it turns out that the approximations do not hurt the classification performance when used for active learning or one-class classification, where only the induced ranking of test examples is of interest. The work was also presented at the NIPS BigVision Workshop in Lake Tahoe, Nevada.

Related Paper:

Rapid Uncertainty Computation with Gaussian Processes and Histogram Intersection Kernels.” Alexander Freytag, Erik Rodner, Paul Bodesheim, and Joachim Denzler. Proceedings of the 11th Asian Conference on Computer Vision, Daejeon, Korea, November 2012.

Add new comment

Plain text

  • No HTML tags allowed.
  • Web page addresses and e-mail addresses turn into links automatically.
  • Lines and paragraphs break automatically.