Abb. INB-Logo

Archiv: Forschung / archive: research


Zurück / back

Multiple motions

 

This page gives a brief overview of our research on multiple motions. The motion patterns that occur in natural scenes are more complex than the state-of-the art motion models. A particular case of more complex motion patterns are multiple overlaid motions that occur in natural environments due to transparencies, reflections, and occlusions.

We first studied multiple overlaid motions in the context of vision science as an extension of research on intrinsic dimension (Zetzsche & Barth, 1990) and visual motion selectivity (Barth & Watson, 2000). Initially, the research on multiple motions was supported by the DAAD grant A/99/22641 to Cicero Mota, who came to visit us from the University of Amazonas in Brazil. We worked out an analytical solution for the problem of estimating transparent motion layers (Mota et al., 2001). The solution is based on a nontrivial linearization of the nonlinear problem. We also performed experiments related to the perception of multiple motion layers (Barth et al., 2001; Mota et al., 2004) and we could predict the ability of human observers to detect the number of motion layers. We argued that detection is limited by the angular separation of the velocities of the layers.

Subsequent research was supported by the DFG - see LOCOMOTOR page - and focused on the mathematical and technical aspects of the problem. We first generalized our approach to include more robust regularization (Stuke et al., 2003) and block-matching (Stuke et al., 2003) techniques. We then extended our analysis to the case of occluded motions (Barth et al., 2003; Mota et al., 2005) and more complex patterns (Scharr et al., 2005). Then we went back to the more basic issue of estimating multiple orientations (Mota et al., 2004; Aach et al., 2006). We then succeeded in solving the more general problem of how to simulteanusly detect and estimate N subspaces (in images this would be orientations and in movies motions) for signals of any dimension (Stuke et al., 2006). This allows us, for example, to estimate an arbitrary number of overlaid orientations in any dimension. Finally, we have shown how to extent the approach to multispectral images (Mota et al., 2006).

 

 

Selected publications


Here we list the publications referenced above. A more comprehensive list of publications can be found on the LOCOMOTOR page and on the more frequently updated INB publications page.

 

     pdf