A hybrid probabilistic model for camera relocalization

Abstract

We present a hybrid deep learning method for modelling the uncertainty of camera relocalization from a single RGB image. The proposed system leverages the discriminative deep image representation from a convolutional neural networks, and uses Gaussian Process regressors to generate the probability distribution of the six degree of freedom (6DoF) camera pose in an end-to-end fashion. This results in a network that can generate uncertainties over its inferences with no need to sample many times. Furthermore we show that our objective based on KL divergence reduces the dependence on the choice of hyperparameters. The results show that compared to the state-of-the-art Bayesian camera relocalization method, our model produces comparable localization uncertainty and improves the system efficiency significantly, without loss of accuracy.