Embedded Files
Abstract
Abstract
We present a method for decomposing the 3D scene flow observed from a moving stereo rig into stationary scene elements and dynamic object motion. Our unsupervised learning framework jointly reasons about the camera motion, optical flow, and 3D motion of moving objects. Three cooperating networks predict stereo matching, camera motion, and residual flow, which represents the flow component due to object motion and not from camera motion. Based on rigid projective geometry, the estimated stereo depth is used to guide the camera motion estimation, and the depth and camera motion are used to guide the residual flow estimation. We also explicitly estimate the 3D scene flow of dynamic objects based on the residual flow and scene depth. Experiments on the KITTI dataset demonstrate the effectiveness of our approach and show that our method outperforms other state-of-the-art algorithms on the optical flow and visual odometry tasks.
Our proposed system learns scene flow and visual odometry with unlabeled stereo videos. Based on the residuals of rigid and non-rigid flow, our system decouples the motions of static and dynamic objects effectively.
Proposed Frameworks
Proposed Frameworks
Experimental Results
Experimental Results
Optical flow estimation on the KITTI scene flow 2015 training set.
Qualitative results of optical flow estimation, which show that the rigid warping errors after the 2nd training phase are significantly refined by the 3rd training phase.
Supplementary Video: Motion Estimation
Supplementary Video: Motion Estimation
Google Sites
Report abuse