Deep Quaternion Pose Proposals for 6D Object Pose Tracking

Abstract

In this work we study quaternion pose distributions for tracking in RGB image sequences the 6D pose of an object selected from a set of objects, for which common models were trained in advance. We propose an unit quaternion representation of the rotational state space for a particle filter, which is then integrated with the particle swarm optimization to shift samples toward local maximas. Owing to k-means++ we better maintain multimodal probability distributions. We train convolutional neural networks to estimate the 2D positions of fiducial points and then to determine PnP-based object pose hypothesis. A CNN is utilized to estimate the positions of fiducial points in order to calculate PnP-based object pose hypothesis. A common Siamese neural network for all objects, which is trained on keypoints from current and previous frame is employed to guide the particles towards predicted pose of the object. Such a key-point based pose hypothesis is injected into the probability distribution that is recursively updated in a Bayesian framework. The 6D object pose tracker is evaluated on Nvidia Jetson AGX Xavier both on synthetic and real sequences of images acquired from a calibrated RGB camera.