Computer Vision News - April 2019

Let’s take a closer look at the LSTM-KF -- long short-term memory Kalman filter, which the authors also define as a model for the temporal regularization of pose estimators. At its core, its idea is to leverage Kalman filters without pre-defining a linear state function A or any constant process, and various measuring matrices Q and R. Instead, the network learns and models a non-linear transition function F, along with matrices Q and R, using 3 different LSTMs - in this way the authors make the method capable of learning a rich and dynamic representation of Kalman filter parameters directly from the data. Architecture Kalman Filter: A quick recap of Kalman Filter equation - is an optimal state estimator if the linearity and Gaussian noise assumptions hold. Where represents the state and our measurement: = −1 + , ∼ (0, ) = + , ∼ (0, ) A , Q , H and R are known matrices The Kalman filter is trained and achieves optimality using an iterative loop with two update steps: the prediction step and the update step. The prediction step estimates the mean and covariance independently of the measurement at time t : ෢′ = ෞ −1 ෢′ = ෢ −1 + Research 6 Research Computer Vision News