TV-L1 optical flow for vector valued images
Research output: Chapter in Book/Report/Conference proceeding › Article in proceedings › Research › peer-review
Standard
TV-L1 optical flow for vector valued images. / Rakêt, Lars Lau; Roholm, Lars; Nielsen, Mads; Lauze, Francois Bernard.
Energy Minimization Methods in Computer Vision and Pattern Recognition: 8th International Conference, EMMCVPR 2011, St. Petersburg, Russia, July 25-27, 2011. Proceedings. ed. / Yuri Boykov; Fredrik Kahl; Victor Lempitsky; Frank R. Schmidt. Springer, 2011. p. 329-343 (Lecture notes in computer science, Vol. 6819).Research output: Chapter in Book/Report/Conference proceeding › Article in proceedings › Research › peer-review
Harvard
APA
Vancouver
Author
Bibtex
}
RIS
TY - GEN
T1 - TV-L1 optical flow for vector valued images
AU - Rakêt, Lars Lau
AU - Roholm, Lars
AU - Nielsen, Mads
AU - Lauze, Francois Bernard
N1 - Conference code: 8
PY - 2011
Y1 - 2011
N2 - The variational TV-L1 framework has become one of the most popular and successful approaches for calculating optical flow. One reason for the popularity is the very appealing properties of the two terms in the energy formulation of the problem, the robust L1-norm of the data fidelity term combined with the total variation (TV) regular- ization that smoothes the flow, but preserve strong discontinuities such as edges. Specifically the approach of Zach et al. [1] has provided a very clean and efficient algorithm for calculating TV-L1 optical flows between grayscale images. In this paper we propose a generalized algorithm that works on vector valued images, by means of a generalized projection step. We give examples of calculations of flows for a number of multi- dimensional constancy assumptions, e.g. gradient and RGB, and show how the developed methodology expands to any kind of vector valued images. The resulting algorithms have the same degree of parallelism as the case of one-dimensional images, and we have produced an efficient GPU implementation, that can take vector valued images with vectors of any dimension. Finally we demonstrate how these algorithms generally produce better flows than the original algorithm.
AB - The variational TV-L1 framework has become one of the most popular and successful approaches for calculating optical flow. One reason for the popularity is the very appealing properties of the two terms in the energy formulation of the problem, the robust L1-norm of the data fidelity term combined with the total variation (TV) regular- ization that smoothes the flow, but preserve strong discontinuities such as edges. Specifically the approach of Zach et al. [1] has provided a very clean and efficient algorithm for calculating TV-L1 optical flows between grayscale images. In this paper we propose a generalized algorithm that works on vector valued images, by means of a generalized projection step. We give examples of calculations of flows for a number of multi- dimensional constancy assumptions, e.g. gradient and RGB, and show how the developed methodology expands to any kind of vector valued images. The resulting algorithms have the same degree of parallelism as the case of one-dimensional images, and we have produced an efficient GPU implementation, that can take vector valued images with vectors of any dimension. Finally we demonstrate how these algorithms generally produce better flows than the original algorithm.
U2 - 10.1007/978-3-642-23094-3_24
DO - 10.1007/978-3-642-23094-3_24
M3 - Article in proceedings
SN - 978-3-642-23093-6
T3 - Lecture notes in computer science
SP - 329
EP - 343
BT - Energy Minimization Methods in Computer Vision and Pattern Recognition
A2 - Boykov, Yuri
A2 - Kahl, Fredrik
A2 - Lempitsky, Victor
A2 - Schmidt, Frank R.
PB - Springer
Y2 - 25 July 2011 through 27 July 2011
ER -
ID: 33478192