Computer Vision News - April 2019

Every month, Computer Vision News reviews a research paper from our field. This month we have chosen Long Short-Term Memory Kalman Filters: Recurrent Neural Estimators for Pose Regularization . We are indebted to the authors ( Huseyin Coskun , Felix Achilles , Robert DiPietro , Nassir Navab , Federico Tombari ), for allowing us to use their images. The paper was first published at ICCV2017 and it is found here . Camera localization, object tracking and pose estimation (for example, localization of joints such as shoulders, knees, hips, wrists, etc. as in illustration) are very changing tasks in computer vision, mainly due to their high dynamic noise. To overcome these tasks/challenges, regularization is often handled using temporal filters: the Kalman filter is the simplest and most general, enjoying widespread use. However, Kalman filters require pre-determining a motion model and measuring model, which raise modeling complexity and are often only a rough approximation of real world behavior. For pose estimation or object tracking, for instance, the motion models used most often assume constant velocity or constant acceleration -- these unrealistic, simplified representations make it very difficult to arrive at a precise solution. The authors propose constructing a network to learn dynamic representations of a motion model and of noise. Specifically, they propose using an LSTM network to learn these models from the data. The proposed method has the advantage of arriving at a representation model which takes account of all previous observations and previous states. In the paper, the authors evaluated their method using the three most popular datasets in this field and achieved state of the art results on each dataset. 4 Research by Assaf Spanier Computer Vision News The authors’ innovation is the LSTM Kalman filter (LSTM-KF), a new architecture capable of learning a motion model and all the parameters of a Kalman filter Research