B04 | Adaptive Algorithms for Motion Estimation

Prof. Andrés Bruhn, Universität Stuttgart
Email | Website

Andres Bruhn

Prof. Dietmar Saupe, Universität Konstanz
Email | Website

Dietmar Saupe

Daniel Maurer, Universität Stuttgart – Email | Website

In contrast to existing approaches that mainly rely on a static model with fixed assumptions, this project aims at designing algorithms that are able to use previously learned knowledge adaptively if applied to novel scenes. This shall be achieved on the basis of a flexible variational model that allows both the integration of different types of prior knowledge and the development of suitable selection mechanisms that seek to identify valid assumptions by a preceding quantitative analysis of the scene. Here, activation weights and model parameters shall be estimated online.

Research Questions

How can typical problems and regularities in image sequences be detected reliably and how can they be quantified numerically?

How can this information be used to rate the difficulty of video sequences, e.g. for assessing benchmarks and for modeling algorithms?

To what extent can priors be learned for different types of problems and regularities?

To what extent can adaptive approaches be designed that are able to handle challenging real-world scenarios based on previously learned priors and decision algorithms?

Adaptive motion estimation. (a) Overlayed input images. (b) Standard approach. (c) Adaptive integration of feature matches at erroneous regions. (d) Final result. Visualization given by color wheel in the center.

Publications

  1. D. Maurer, M. Stoll, S. Volz, P. Gairing, and A. Bruhn, “A comparison of isotropic and anisotropic second order regularisers for optical flow.,” in International Conference on Scale Space and Variational Methods in Computer Vision (SSVM)., Berlin, 2017, vol. Lecture Notes in Computer Science, no. 10302, pp. 537–549.
  2. D. Maurer, M. Stoll, and A. Bruhn, “Order-adaptive and illumination-aware variational optical flow refinement,” in British Machine Vision Conference (BMVC), 2017.
  3. M. Stoll, S. Volz, D. Maurer, and A. Bruhn, “A time-efficient optimisation framework for parameters of optical flow methods,” in Scandinavian Conference on Image Analysis (SCIA)., Berlin, 2017, vol. Lecture Notes in Computer Science, no. 10269, pp. 41–53.
  4. K. Kurzhals, M. Stoll, A. Bruhn, and D. Weiskopf, “FlowBrush: Optical Flow Art,” in Symposium on Computational Aesthetics, Sketch-Based Interfaces and Modeling, and Non-Photorealistic Animation and Rendering (EXPRESSIVE, co-located with SIGGRAPH)., 2017.
  5. D. Maurer, M. Stoll, and A. Bruhn, “Order-adaptive regularisation for variational optical flow: global, local and in between,” in International Conference on Scale Space and Variational Methods in Computer Vision (SSVM)., Berlin, 2017, vol. Lecture Notes in Computer Science, no. 10302, pp. 550–562.