Output details
11 - Computer Science and Informatics
University College London
Learning a Confidence Measure for Optical Flow
<23>This paper established that a supervised learner can be trained to pick which of several "expert" algorithms will score highest for each variant of a problem. This is a modern and practical twist on example-based cost-sensitive learning. This and our 2010 conference paper validated this Bayesian model across three different Holy-Grail problems in Computer Vision: optical flow, descriptor matching, and patch tracking. Significance: Now being evaluated for lung-cancer screening at UCL Centre for Respiratory Research. Invited to give Rank Prize Symposium talk because this method discovers great specialist algorithms (that may be poor on average) and when to use them.