Complementary Optic Flow on GPU - Mathematical Image Analysis Group, Saarland University

Complementary Optic Flow on the GPU

Model
Complementary Optic Flow

Our highly efficient GPU algorithm is based on the Complementary Optic Flow [1] model by Zimmer et al. According to the Middlebury benchmark, it is currently one of the most accurate optic flow algorithms. Moreover, it is formulated in a general variational framework which entails that it is flexible and can easily be extended to new features in the future.

The key idea behind the Complementary Optic Flow model is its anisotropic regulariser which steers the smoothing in accordance to the constraints imposed in the data term. As a consequence, data and smoothness term act ‘complementary’, i.e. they support each other instead of interfering. Such a behaviour is achieved by designing the anisotropic regulariser in a way that it only smoothes across feature gradients, but not along them. Thus, the smoothness term still successfully remedies the aperture problem (fills in information where the data term gives no information), but does not smooth away important flow features.

Since we are interested in a small runtime, we slightly modified the original model, and introduced tradeoffs between speed and accuracy. To this end, we resort to the standard RGB colour space instead of using a HSV colour representation, we propose a new resampling strategy that profits from hardware acceleration by the graphics card, and we use a different solver (FED solver) which offers the opportunity to control the desired stopping time of the process more accurately in favour of a smaller runtime.

References:

[1] H. Zimmer, A. Bruhn, J. Weickert, L. Valgaerts,
A. Salgado, B. Rosenhahn, H.P. Seidel:
Complementary optic flow.
In D. Cremers, Y. Boykov, A. Blake, F. R. Schmidt (Eds.): Energy Minimization Methods in Computer Vision and Pattern Recognition (EMMCVPR). Lecture Notes in Computer Science, Vol. 5681, 207-220, Springer, Berlin, 2009.

Main Page FED >