Master Thesis MSTR-2020-33

BibliographyMehl, Lukas Francesco: Anisotropic selection schemes for order-adaptive variational optical flow methods.
University of Stuttgart, Faculty of Computer Science, Electrical Engineering, and Information Technology, Master Thesis No. 33 (2020).
86 pages, english.
Abstract

Variational optical flow estimation, i.e. the estimation of motion information from image sequences using continuous optimization approaches, plays an important role in computer vision. Recently, the order-adaptive regularization approach by Maurer et al. [MSB17b] has shown to yield a performance improvement by not restricting itself to first- or second-order regularization but selecting the optimal order locally adaptive. Although the regularizers included in their method are anisotropic, i.e. adaptive to local image directions, the order selection is not. The selection scheme is restricted to always choosing the same order for different local directions, which is a shortcoming that will be addressed in this thesis. After providing an introduction into the topic and a detailed derivation, minimization and discretization of the method by Maurer et al., this thesis presents the model for an order-adaptive approach that includes an anisotropic order selection scheme. The newly presented method selects the regularization order locally adaptive and individually for two directions corresponding to image structures. Additionally, by deriving the minimization of the novel method, it is shown that the same optimization as in the work of Maurer et al. [MSB17b] is possible. Finally, the approaches are compared using an evaluation on recent benchmarks and an analysis of the resulting order maps is given. While the results do not show an improvement on the benchmarks, experiments including the ground truth data show that the idea has potential.

Full text and
other links
Volltext
Department(s)University of Stuttgart, Institute of Visualisation and Interactive Systems, Visualisation and Interactive Systems
Superviser(s)Bruhn, Prof. Andres; Maurer, Dr. Daniel; Stoll, Michael
Entry dateDecember 17, 2020
   Publ. Computer Science