Shape from rotation

1991 Computer Vision and Pattern Recognition |

Published by IEEE


The construction of a 3D surface model of an object rotating in front of a camera is examined. Previous research in depth from motion has demonstrated the power of using an incremental approach to depth estimation. The author extends this approach to more general motion and uses a full 3D surface model instead of a 2 1/2 D depth map. The algorithm starts with a flow field computed using local correlation. It then projects individual measurements into 3D points with associated uncertainties. Nearby points from successive frames are merged to improve the position estimates. These points are then used to construct a deformable surface model, which is refined over time. The application of novel techniques to several image sequences is demonstrated. >