Triangulation-Based Approaches to Three-Dimensional Scene Reconstruction - 3D Computer Vision: Efficient Methods and Applications - page 68

Fig. 1.22 Determination of
scene flow information
(adapted from Huguet and
Devernay, 2007 )
field over two stereo image pairs acquired at subsequent time steps, Wedel et al.
( 2008a ) propose a separate determination of disparity and optical flow while keep-
ing both estimates consistent with each other. At the same time, individual methods
which are optimal e.g. with respect to accuracy or computational efficiency can be
employed for each of these two parts of the scene flow estimation problem. The
epipolar constraint is exploited for the computation of disparity and the vertical
component of the optical flow. An energy functional consisting of the weighted sum
of a first term describing the difference between the grey values in the left and the
right image at the current and the subsequent time step, respectively ('data term'),
and a second term favouring small absolute values of the gradients of the optical flow
components and the disparity ('smoothness term') is minimised by a variational ap-
proach. An extensive experimental evaluation is performed by Wedel et al. ( 2008b ),
regarding image sequences synthetically generated with a ray tracing software, such
that ground truth data are available, and real-world urban traffic scenes. Wedel et al.
( 2011 ) compare several disparity estimation approaches in the context of their scene
flow computation framework, namely the sparse methods by Franke and Joos ( 2000 )
and by Stein ( 2004 ) as well as the semi-global matching approach based on mutual
information by Hirschmüller ( 2008 ), where the best results are obtained using the
method by Hirschmüller ( 2008 ). An experimental evaluation regarding synthetically
generated image sequences and real-world traffic scenes, including long stereo im-
age sequences, is presented by Wedel et al. ( 2011 ). The mean errors of the estimated
optical flow components and the disparity values are typically well below 1 pixel,
with standard deviations ranging from some tenths of a pixel to about 2 pixels. Fur-
thermore, the scene flow field is used by Wedel et al. ( 2011 ) for extracting objects
from the scene which move in a manner different from the background, relying on
a graph cut-based segmentation approach.
Local Intensity Modelling In contrast to the previously described spacetime
stereo approaches, which exploit a pixel-based similarity measure, the spacetime