CeMNet: Self-Supervised Learning for Accurate Continuous Ego-Motion Estimation

In this paper, we propose a novel self-supervised learning model for\nestimating continuous ego-motion from video. Our model learns to estimate\ncamera motion by watching RGBD or RGB video streams and determining\ntranslational and rotation velocities that correctly predict the appearance of\nfuture frames. Our approach differs from other recent work on self-supervised\nstructure-from-motion in its use of a continuous motion formulation and\nrepresentation of rigid motion fields rather than direct prediction of camera\nparameters. To make estimation robust in dynamic environments with multiple\nmoving objects, we introduce a simple two-component segmentation process that\nisolates the rigid background environment from dynamic scene elements. We\ndemonstrate state-of-the-art accuracy of the self-trained model on several\nbenchmark ego-motion datasets and highlight the ability of the model to provide\nsuperior rotational accuracy and handling of non-rigid scene motions.\n