Information Technology Reference
In-Depth Information
Photometric variations : it is quite reasonable to assume that homologous points
in the left and right views have identical intensity values. This assumption, com-
monly referred to as the brightness constancy assumption, is sometimes violated
in practice due to varying shadows, illumination changes and specular reflections.
Thus, although commonly used, this hypothesis may lead to incorrect disparity
estimates and consequently may reduce the efficiency of depth recovery.
Untextured regions : image areas which contain little or repetitive textures result
in ambiguities in the matching process caused by the presence of multiple possi-
ble matches of very similar intensity patterns.
Depth discontinuities : the presence of discontinuities causes occlusions, which
are points only visible in one image of the stereo pair, making the disparity as-
signment very difficult at object boundaries.
In order to overcome these ambiguities and make the problem more tractable, a
variety of constraints and assumptions are typically made. The most commonly used
constraints are related to the following factors:
Smoothness : this assumption imposes a continuous and smooth variation in the
uniform areas of the disparity field. It is motivated by the observation that natural
scenes consist of objects with smooth surfaces. It holds true almost everywhere
except at depth boundaries.
Uniqueness : states that a pixel of one view can have, at most, one corresponding
pixel in the other view. This constraint is often used to identify occlusions by
enforcing one-to-one correspondences for visible pixels across images.
Ordering : constrains the order of points along epipolar lines to remain the same.
The advantage of using this assumption is that its application allows for the ex-
plicit detection of occlusions. However, it does not always hold true, especially
for scenes containing thin foreground objects.
2.3.1
A Survey of Stereo Vision Methods
There is a considerable amount of literature on the stereo correspondence problem.
An extensive review is addressed by Scharstein and Szeliski in [4]. The authors
identify four steps that characterize stereo algorithms. These are:
(i) matching cost computation,
(ii) cost aggregation,
(iii) optimization,
(iv) disparity refinement.
In the analysis below, we focus on the optimization component and classify stereo
algorithms as local, progressive, cooperative or global optimization methods.
 
Search WWH ::




Custom Search