Page 1 of 10
Journal for Studies in Management and Planning
Available at
http://edupediapublications.org/journals/index.php/JSMaP/
e-ISSN: 2395-0463
Volume 02 Issue 12
December 2016
Available online: http://edupediapublications.org/journals/index.php/JSMaP/ P a g e | 92
Continuous Depth Map Reconstruction from Light Fields
Meghavath Kavitha, & Mrs.G.Satya Prabha
1PG Scholar, Department of ECE, SLC's Institute of Engineering and Technology, Piglipur Village, Hayathnagar
Mandal, Near Ramoji Film City, Ranga Reddy District, Hyderabad, Telangana
2Assistant Professor, Department of ECE, SLC's Institute of Engineering and Technology, Piglipur Village,
Hayathnagar Mandal, Near Ramoji Film City, Ranga Reddy District, Hyderabad, Telangana
Abstract In this paper, we investigate how
the recently emerged photography
technology—the light field—can benefit
depth map estimation, a challenging
computer vision problem. A novel
framework is proposed to reconstruct
continuous depth maps from light field data.
Unlike many traditional methods for the
stereo matching problem, the proposed
method does not need to quantize the depth
range. By making use of the structure
information amongst the densely sampled
views in light field data, we can obtain
dense and relatively reliable local
estimations. Starting from initial
estimations, we go on to propose an
optimization method based on solving a
sparse linear system iteratively with a
conjugate gradient method. Two different
affinity matrices for the linear system are
employed to balance the efficiency and
quality of the optimization. Then, a depth- assisted segmentation method is introduced
so that different segments can employ
different affinity matrices. Experiment
results on both synthetic and real light fields
demonstrate that our continuous results are
more accurate, efficient, and able to
preserve more details compared with
discrete approaches.
I. INTRODUCTION
Depth map reconstruction, also known as
disparity estimation, is a traditional
challenging computer vision task which has
been studied for more than three decades.
The conventional way to get depth values is
from stereo images. Scharstein and Szeliski
gave a good survey of this topic in Most
methods nowadays solve the problem by
minimizing an energy function, which
usually consists of a data term and a
smoothness term. The two most popular
models for the energy functions are the
Markov Random Fields model (MRFs) and
variational approaches while the formulation
of the data term and the smoothness term
can be different. For stereo matching, a lot
of previous work shows that energy
Page 2 of 10
Journal for Studies in Management and Planning
Available at
http://edupediapublications.org/journals/index.php/JSMaP/
e-ISSN: 2395-0463
Volume 02 Issue 12
December 2016
Available online: http://edupediapublications.org/journals/index.php/JSMaP/ P a g e | 93
functions with non-convex terms model the
problem better, although in fact only an
approximate optimization can be found. In
this case, the recovered depth values are not
continuous but discrete in the depth range. A
drawback of the discrete methods is that the
time and memory cost of the algorithm is
related to the number of quantized levels.
When the level is not high
The newly developed technology the light
field has brought new possibilities in
reconstruction of depth maps. Compared
with traditional image data, a light field
contains not only accumulated colour
intensities but also some information about
ray directions. In general, the light field data
can be seen as a set of photos captured from
densely and regularly placed cameras. When
the views are dense enough, there are some
interesting properties that make light fields
different from traditional multi-view data.
One of the properties that has been studied
to solve traditional computer vision
problems is the fact that the projected points
from one 3D point onto different views
correspond to a line in the so called
epipolarplane image (EPI). The slope of this
line is related to the depth of the point in the
space, which transforms the problem of
depth estimation into line detection on EPIs.
To estimate the orientation of the lines on
EPIs, one idea is to try out all different
orientations: the one with the least colour
variance along the line is most likely to give
the correct depth value. Several methods
have been developed based on this point;
different methods use different ways to
measure the colour variance. Kim et al.
employed a modified
Parzen window estimation with an
Epanechenikov kernel Tao et al., on the
other hand, used the standard deviation
estimation. Defocus is another clue that can
be used for depth. Instead of using the
colour variance, researchers try to find out
by how much angle the EPIs need to be
sheared to make the interest point in focus.
Defocus and colour variance are used
together to find the depth . An alternative
approach proposed by Wanner and
Goldluecke is to use a structure tensor to
estimate the slope of the lines on EPIs.
Unlike the other methods, it does not need to
try out different hypothetical depth values to
find the optimal one, but at once provides an
estimation as well as a certainty level from
one structure tensor operation. In this work,
we use the estimation from the structure
tensor as a starting point, followed by a
refinement step by examining the colour
Page 3 of 10
Journal for Studies in Management and Planning
Available at
http://edupediapublications.org/journals/index.php/JSMaP/
e-ISSN: 2395-0463
Volume 02 Issue 12
December 2016
Available online: http://edupediapublications.org/journals/index.php/JSMaP/ P a g e | 94
correspondence along the detected line from
the structure tensor. The initial estimations
from light fields are denser and more
reliable compared with those from
traditional stereo images, which also makes
the optimization step different. However,
some works still follow a similar
optimization method as with stereo
matching, such as MRFs in and functional
lifting in .These methods have to discretize
the depth values, so lose the advantages of
the dense and reliable initial estimations.
Wanner and Goldluecke showed in
Fig. 1. Two-plane parametrization of 4D
light fields. that a simple denoising filter can
generate comparable results with the
discrete global optimization,
since the filter keeps the continuous depth
values. In Kim et al.’s work by applying the
local method iteratively on the EPIs with
different resolutions, global estimation can
be avoided. Only a median filter is applied
to eliminate outliers. With the
aforementioned initial estimation methods,
the depth values usually come with certainty
levels. Therefore the optimization problem
is in essence to propagate the reliable
estimations to other parts. In this work, we
explore an optimization method, given by
solving a linear system, which generates a
smooth and globally optimized result.
II. RELATED WORK
The light field concept was originally
defined by physicists who interpreted the
flow of light as a field. It is a plenoptic
function which describes the amount of
light, also known as radiance, travelling
towards every direction through every point
in the space. However, to measure the
radiance of the light at every location
towards every direction is not feasible in
practice. The capture of light fields in fact
relies on sampling the radiance in space and
reconstructing the plenoptic function. The
4D light field was first proposed and later
widely used in light field analysis. We adopt
the two-plane parametrization of 4D light
fields and denote it as L(x, y, s, t), as shown
in Figure 1. Under this parametrization, a
4D light field can be seen as a 2D array of
perspective views, where (s, t) can be seen
as the index of different views and (x, y) are
spatial coordinates within each view
C:\Users\Admin\Desktop\f1.png
