Unsupervised learning for monocular dynamic motion segmentation

Abstract

Dynamic object motion estimation is important for robotic and automatic driving. However, the method of object motion not only depend on manual parameter adjustment，but also need large motion object label which is difficult achieved. In this paper, we propose a new dynamic rigid object motion segmentation frame, which we combine optical flow, depth map, 6-DoF pose estimation with motion segmentation. Especially, we utilize the optical flow in pose estimation and supervising the depth estimation. In the experiment, we evaluate our frame on KITTI Scenes Flow dataset. The result show that our method could accurately estimate motion object.