fast pose estimation with parameter sensitive hashing.pdfVIP

Fast Pose Estimation with Parameter Sensitive Hashing Gregory Shakhnarovich Paul Viola Trevor Darrell gregory@ viola@ trevor@ Computer Science and Artificial Intelligence Lab, MIT Microsoft Research Cambridge, MA 02139 Redmond, WA 98052 Abstract assumed to be sufficiently smooth to make such inference meaningful. Example-based methods are effective for par ameter es- Classic methods for example-based learning, such as the timation pr oblems when the underlying system is simple -nearest neighbor rule (-NN) and locally-weighted re- or the dimensionality of the input is low. For complex gression (LWR), are appealing due to their simplicity and and high-dimensional pr oblems such as pose estimation, the asymptotic optimality of the resulting estimators. How- the number of required examples and the computational ever, the computational complexity of similarity search in complexity rapidly become pr ohibitively high. We intro- high-dimensional spaces and on very large data sets has duce a new algorithm that learns a set of hashingfunctions made these methods infeasible for many vision problems. that efficiently index examples relevant to a particular es- In this paper we describe a new example-based algorithm timation task. Our algorithm extends a recently developed for fast parameter estimation using local models, which