Unsupervised Joint Multi-Task Learning of Vision Geometry Tasks

In this paper, we present a novel architecture and training methodology for learning monocular depth prediction, camera pose estimation, optical flow, and moving object segmentation using a common encoder in an unsupervised fashion. We demonstrate that the geometrical relationships between these tasks not only support joint unsupervised learning as shown in previous works but also allow them to share common features. We also show the advantage of using a two-stage learning approach to improve the performance of the base network.