Abstract
Conventional image motion based structure from motion methods first computeoptical flow, then solve for the 3D motion parameters based on the epipolarconstraint, and finally recover the 3D geometry of the scene. However, errorsin optical flow due to regularization can lead to large errors in 3D motion andstructure. This paper investigates whether performance and consistency can beimproved by avoiding optical flow estimation in the early stages of thestructure from motion pipeline, and it proposes a new direct method based onimage gradients (normal flow) only. The main idea lies in a reformulation ofthe positive-depth constraint, which allows the use of well-known minimizationtechniques to solve for 3D motion. The 3D motion estimate is then refined andstructure estimated adding a regularization based on depth. Experimentalcomparisons on standard synthetic datasets and the real-world driving benchmarkdataset KITTI using three different optic flow algorithms show that the methodachieves better accuracy in all but one case. Furthermore, it outperformsexisting normal flow based 3D motion estimation techniques. Finally, therecovered 3D geometry is shown to be also very accurate.