Exploiting uncertainty in regression forests for accurate camera relocalization

Julien Valentin, Matthias Niebner, Jamie Shotton, Andrew Fitzgibbon, Shahram Izadi, Philip Torr
2015 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)  
Recent advances in camera relocalization use predictions from a regression forest to guide the camera pose optimization procedure. In these methods, each tree associates one pixel with a point in the scene's 3D world coordinate frame. In previous work, these predictions were point estimates and the subsequent camera pose optimization implicitly assumed an isotropic distribution of these estimates. In this paper, we train a regression forest to predict mixtures of anisotropic 3D Gaussians and
more » ... w how the predicted uncertainties can be taken into account for continuous pose optimization. Experiments show that our proposed method is able to relocalize up to 40% more frames than the state of the art. The main contributions of this work are (i) the extension of the state of the art on RGB-D camera relocalization by modeling and minimizing uncertainties for regression tree induction and predictions performed by the regression forest; and (ii) leveraging these uncertainties in order to provide for improved relocalization without using explicit models of the scenes. The proposed camera relocalization approach consists of two major components: (i) a regression forest trained on RGB-D input data to predict anisotropic Gaussian mixtures of 3D scene coordinates; and (ii) a continuous pose optimization leveraging the anisotropic Gaussian mixtures predicted by the forest. An overview of the complete relocalization pipeline is shown in Fig. 1 .
doi:10.1109/cvpr.2015.7299069 dblp:conf/cvpr/ValentinNSFIT15 fatcat:rihvn3n4jjfebaihlccqif7wqq