An Unbiased Second-Order Prior for High-Accuracy Motion Estimation

Werner Trobin, Thomas Pock, Daniel Cremers, Horst Bischof

Abstract

Virtually all variational methods for motion estimation regularize the gradient of the flow field, which introduces a bias towards piecewise constant motions in weakly textured areas. We propose a novel regularization approach, based on decorrelated second-order derivatives, that does not suffer from this shortcoming. We then derive an efficient numerical scheme to solve the new model using projected gradient descent. A comparison to a TV regularized model shows that the proposed second-order prior exhibits superior performance, in particular in low-textured areas (where the prior becomes important). Finally, we show that the proposed model yields state-of-the-art results on the Middlebury optical flow database.
BibTeX
@inproceedings{trobin2008unbiased,
  author = {Werner Trobin and Thomas Pock and Daniel Cremers and Horst Bischof},
  title = {{An Unbiased Second-Order Prior for High-Accuracy Motion Estimation}},
  booktitle = {DAGM-Symposium},
  year = {2008},
  doi = {10.1007/978-3-540-69321-5\_40},
}

Loading related content...

Figure