Three-Dimensional Pose Estimation and Segmentation Methods - 3D Computer Vision: Efficient Methods and Applications - page 115

Fig. 2.10 Extraction and
projection of the
three-dimensional contour
model for camera 1

single step of a Newton-Raphson optimisation procedure. The learned probability
distributions S c ( m T ,Σ T ) are used to maximise the joint probability

N c

p T i 1 ,...,I N c = p I c | S c ( m T ,Σ T ) m T , Σ T )

| · p( T | (2.30)

c = 1

with S c ( m T ,Σ T ) representing the probability distributions close to the pro-
jected curve in image I c . As in the original CCD framework, we optimise
the log-likelihood of ( 2.30 ) with a Newton-Raphson optimisation step. The
MOCCD algorithm can be illustrated as follows: The Gaussian probability den-
sities p(I ∗ |

S c ( m T ,Σ T )) are sensitive with respect to step-like structures along the
curve perpendiculars in each image I c , and the three-dimensional model is adapted
to the N c camera images by an implicit triangulation.

2.2.3.4 The Shape Flow Algorithm

The shape flow (SF) algorithm is introduced by Hahn et al. ( 2008b , 2010a )asatop-
down approach for spatio-temporal pose estimation. In contrast to bottom-up motion
estimation approaches, like the motion analysis module described in Sect. 2.3.3 ,the
SF algorithm is generally able to estimate the three-dimensional pose parameters T
and the temporal pose derivative

T with a spatio-temporal model and the images of