SKen: A Statistical Test for Removing Outliers in Optical Flow
A 3D Reconstruction Case
Samuel Macedo, Luis Vasconcelos, Vincius Cesar, Saulo Pessoa and Judith Kelner
GRVM, CIn (UFPE), Recife, Brazil
Keywords:
Outlier Detection, Optical Flow, Computer Vision, 3D Reconstruction, Statistical Inference.
Abstract:
The 3D reconstruction can be employed in several areas such as markerless augmented reality, manipulation
of interactive virtual objects and to deal with the occlusion of virtual objects by real ones. However, many
improvements into the 3D reconstruction pipeline in order to increase its efficiency may still be done. In such
context, this paper proposes a filter for optimizing a 3D reconstruction pipeline. It is presented the SKen
technique, a statistical hypothesis test that classifies the features by checking the smoothness of its trajectory.
Although it was not mathematically proven that inliers features performed smooth camera paths, this work
shows some evidence of a relationship between smoothness and inliers. By removing features that did not
present smooth paths, the quality of the 3D reconstruction was enhanced.
1 INTRODUCTION
Computer vision is a research area with numerous
contributions to the development of 3D reconstruc-
tion techniques and it is mainly concerned with the
modeling of real world information (Hming and Pe-
ters, 2010) (Barbosa, 2006). Advances in this area
involve the integration with algorithms for real-time
execution (Nist
´
er, 2003), robust statistical approaches
(Choi and Medioni, 2009) and dense 3D reconstruc-
tion hardware-based acceleration methods (Bouguet,
2000). This way, several challenges still motivate the
development of new techniques and the interoperabil-
ity of these with the existing ones.
When the 3D reconstruction is made from real
data, there is an introduction of errors correspondent
to the quality of the image acquisition. This procedure
depends on parameters such as the image resolution,
the camera sensor and light conditions. Once the pro-
cessing of this image already presents noise, it will
be passed on to the next steps. This way, these errors
cause more errors to accumulate along the pipeline.
Following, in the tracking phase there are also
some inherent difficulties such as the occlusion of part
of the scene and false matching. This is due to the
image areas with poor textures and gradient with low
significance that may also introduce noise in the posi-
tioning of features along the trace.
Considering a priori that there are errors aris-
ing from both the image acquisition and the track-
ing stage, it is essential an approach takes into ac-
count these errors when calculating the matrix camera
and the fundamental matrix (Hartley and Zisserman,
2004).
An optimal reconstruction happens when the data
from the tracking stage were obtained without errors,
i.e. using the ground truth. Since the actual distri-
bution of the errors is unknown, it is not possible, at
first, to filter in the scene which are the reliable fea-
tures (inliers) and the unreliable ones (outliers).
Numerical errors, the minimization of nonlinear
systems and several other limitations of the pipeline
make it extremely difficult to avoid the accumulation
of past errors from one phase to another. A solution
to this problem would reduce the amount of errors in
the first stage of the pipeline to its minimum, thereby
providing the best possible features to the 3D recon-
struction algorithm.
In this context, this paper proposes a methodol-
ogy based on statistical methods in order to filter fea-
tures during the tracking phase of a 3D reconstruc-
tion pipeline, using only 2D points provided by the
tracker. It was developed the SKen, a robust hypoth-
esis test with low computational cost, which aims to
reduce the error caused by the tracker without affect-
ing the total execution time.
The remaining sections of this article are orga-
nized as follows: section 2 contemplates some tech-
niques on the state of the art, section 3 describes the
SKen test, section 4 comprises the validation of the
202
Macedo S., Vasconcelos L., Cesar V., Pessoa S. and Kelner J..
SKen: A Statistical Test for Removing Outliers in Optical Flow - A 3D Reconstruction Case.
DOI: 10.5220/0004748802020209
In Proceedings of the 9th International Conference on Computer Vision Theory and Applications (VISAPP-2014), pages 202-209
ISBN: 978-989-758-003-1
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
test in both synthetic and real scenarios, and finally,
section 5 discusses the advantages and some sugges-
tions for improving the methodology.
2 OUTLIER DETECTION
Once the information extracted from real world is al-
ready noisy, the 3D reconstruction calculations will
not generate flawless results, but a hypothesis instead
(an intermediate product). A useful method for eval-
uating hypotheses is the RANSAC algorithm (Fis-
chler and Bolles, 1987), which consists of an itera-
tive method to estimate parameters of a mathematical
model based on a set of observations that contains er-
rors.
The RANSAC performs two steps as follows. A
hypothesis is randomly selected and tested with the
full universe of data; if the tests confirm the hypoth-
esis through a threshold determined by the user, this
assumption is saved as a candidate to the final prod-
uct. After estimating several hypotheses, the best one
is chosen as the final product according to predeter-
mined parameters in the process.
The observations that are consistent with the
mathematical model are named inliers and those that
do not meet the predetermined parameters are consid-
ered outliers. To determine whether an observation is
inlier or outlier, the error the hypotheses generate is
compared to a threshold determined by the user. The
RANSAC chooses as its best hypothesis the one with
the highest number of inliers.
A disadvantage of RANSAC is that there is no up-
per bound on the time it takes to compute these pa-
rameters. When the number of performed iterations
is limited, the obtained solution may not be optimal;
in fact, it may not even be minimally appropriate for
the data. In this way, the RANSAC offers the follow-
ing trade-off: a greater number of iterations increases
the probability of conceiving a reasonable model, al-
though the total execution time is also increased. A
final disadvantage of RANSAC is that it requires the
setting of problem-specific thresholds.
Currently, there are many algorithms that are ad-
justments or enhancements of the RANSAC: LMedS
(Rousseeuw and Leroy, 1987), GASAC (Rodehorst
and Hellwich, 2006), StaRSaC (Choi and Medioni,
2009), MSAC (Torr and Zisserman, 2000) and MLE-
SAC (Torr and Zisserman, 2000). These algorithms
are supposed to be more robust than RANSAC. Nev-
ertheless, they all present the same drawback: higher
computational cost. If the input data could have few
outliers initially, the RANSAC alone would generate
good hypothesis in the shortest time possible.
In such context, it would be ideal to remove out-
liers between the tracking phase and the RANSAC
execution. Thus, with as few outliers as possible,
less iterations are necessary and the pipeline can im-
prove its the performance. To achieve that, a tech-
nique for removing outliers should have a computa-
tion cost which is practically imperceptible.
3 SKen TEST
Although not formally proven until the present date,
it is common sense that inliers features follow smooth
paths while outliers do not present smoothness in a
camera path. It was not found in the literature a test
or technique with low computational cost whose pur-
pose was to quantify the smoothness for a camera
path. Thus, in this paper it is presented a hypothesis
test capable of evaluating the smoothness of the fea-
ture path. The hypothesis test proposed, named SKen,
was applied in the context of optical flow, for the paths
of features tracked in the scene. The technique pre-
sented in this paper proposes a methodology for iden-
tifying and ranking features in order of smoothness.
An important factor of the proposed methodology is
that users do not have to enter any parameters in order
to execute the SKen. The result is an automatic and
deterministic method.
3.1 Random Variable
Suppose two features a and b in a video sequence.
At ten frames, the feature a moved to the coordinates
C
a
=(1, 8, 12, 14, 13, 8, 4, 5, 8, 12) and the feature b
moved to C
b
=(1, 8, 12, 10, 13, 8, 9, 5, 8, 12) resulting
in two paths which can be seen in figure 1.
Figure 1: (a) Path of feature a (b) Path of feature b.
The feature a presents a smooth behavior whereas
the feature b is clearly noisy. To verify the smooth-
ness, it is used the second derivative of the function
that generated the feature path. Once the function is
discrete, the derivatives have to be approximated. The
derivatives values are shown in table 1.
It can be observed that for the a feature, the sec-
ond derivative alternate the signal only once due to
SKen:AStatisticalTestforRemovingOutliersinOpticalFlow-A3DReconstructionCase
203
Table 1: Values for approximated derivatives of 2nd order
for the features a and b.
f
00
(a) -3 -2 -3 -4 1 5 2 1
f
00
(b) -3 -6 5 -8 6 -5 7 1
the inflection point, whereas for feature b the second
derivative switches 5 times.
In summary, there was only one alternation of sig-
nal into eight values of f
00
(a) against five alternations
of signals for the same of values of f
00
(b). It is neces-
sary a hypothesis test in order to evaluate how many
alternations of signs are valid to assure that a path is,
indeed, smooth. Thus, the alternating signal of the
second derivative is the random variable under inves-
tigation. It was applied the nonparametric test SKen
for the random variable K =
i
k
i
so that:
k
i
=
1, if f
00
[i + 1] f
00
[i] > 0
1, if f
00
[i + 1] f
00
[i] < 0
0, c.c.
(1)
where f
00
[i] is the sequence of second derivatives of
the path in question.
3.2 Assumptions
Like in all statistical hypothesis tests, it is necessary
to define the assumptions, which must be satisfied in
order to guarantee proper precision to the test as well
as avoiding entailing wrong decisions. They are:
ensure that the sequence is derived from a path
the data must be in temporal order
the sequence cannot be extremely sinuous
In case a sequence is extremely sinuous, the
amount of inflection points may impair the test by
making it more rigorous than necessary. In a long
and not excessively sinuous sequence, the number of
inflection points can be ignored.
3.3 The SKen Statistics
As defined in section 3.1, the random variable in ques-
tion is the sum of the number of signal transition in the
values of second derivatives (K =
i
k
i
). Considering
the example given in section 3.1, the value of K for
the feature a and b is illustrated in table 2.
It can be noticed that the feature a (
i
k
i
(a) = 5)
Table 2: Values of k
i
and K for the features a and b.
k
i
(a) +1 +1 +1 -1 +1 +1 +1
k
i
(b) +1 -1 -1 -1 -1 -1 +1
has a larger K than the feature b (
i
k
i
(a) = 3), i.e.
the feature b has more alternating signs that the fea-
ture a, hence a is smoother than b. In other words: the
higher the K, the smoother the path.
If the path has n terms, the maximum value of K is
K = n 3 and the minimum value is K = (n3), so
the path needs to have at least four terms. The exact
probability distribution of K for n terms is presented
in Table 3:
Table 3: Exact probability distribution of the variable K for
n terms.
K -(n-3) -(n-5) . . . (n-5) (n-3)
P(K = k
i
)
(
n
0
)
2
n3
(
n
1
)
2
n3
. . .
(
n
n1
)
2
n3
(
n
n
)
2
n3
3.4 Large-sample Approximation
When n increases the calculation of n! becomes com-
putationally costly and this is out of the outline of this
paper, so it has to be checked an approximation to the
probability distribution P(K = k
i
). The central limit
theorem (James, 2002) treats the convergence in dis-
tribution of normalized partial sums for the standard
normal distribution N(0.1). This way, it is assumed
that all variances are finite and that at least one of
them is strictly positive. The problem lies in finding
conditions under which:
S
n
E(S
n
)
p
Var(S
n
)
D
N(0, 1) (2)
In such way, it is necessary to calculate E(K) and
Var(K). Assuming a feature with a path containing n
terms, so n3 transitions, the calculation of E(K) and
Var(K) are:
E(K) =
(n 3)
n
0
+ ···+(n 3)
n
n
2
n3
= 0 (3)
Var(K) =
(n 3)
2
n
0
+ ···+(n 3)
2
n
n
2
n3
= n 3
(4)
When using a continuous approximation is at-
tributed to a discrete variable, it is expected that some
adjustment must be made. This adjustment is called
continuity correction (Bussab and Morettin, 2002).
The correction is performed by subtracting 1 to the
value of K before calculating the value of the normal.
Thus, given a feature with n terms, the approximation
of the probability distribution of K to the Gaussian
distribution is:
VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications
204
K 1
n 3
D
N(0, 1) (5)
It must be highlighted that, due to the characteris-
tic of the equation 5 itself, features with longer paths
are considered smoother than features with shorter
paths. Whenever n increases, the numerator increases
faster than the denominator. For example, suppose a
feature with ten coordinates and K equal to ten and
another one with twenty coordinates and K equals
twenty. Both have all the second derivatives with
the same sign, indicating the maximum smoothness.
However, the p-value of the first, with ten coordinates,
is 3x10
4
and the second, with twenty coordinates,
would have p-value 2x10
6
. Thus the feature with
the longest path would be placed first in the order of
smoothness than the one with shorter.
3.5 SKen Test Application
The methodology for applying the SKen test is given
by the algorithm 1.
Algorithm 1: SKen test algorithm.
1. Track the scene using a generic tracker.
2. Decompose the input given by the tracker in
(X,Y). That is, for every feature we have the path it
does in X and the path it does in Y.
3. Apply the SKen test in all feature for both co-
ordinates (X,Y) and sort the paths by p-values in
ascending order, i.e. in order of smoothness.
4. Verify the requirements of the 3D reconstruction
algorithm.
5. Provide as input to the 3D reconstruction algo-
rithm the smoother features that fulfill its require-
ments. A feature is only considered smooth when
is smooth on X and Y.
If the video to be reconstructed is filmed predomi-
nantly on a single axis, either X or Y, the methodology
can be executed by applying the SKen test only to this
axis and thereby decreasing the execution time.
4 RESULTS
This section presents the main results obtained by us-
ing the methodology proposed in this paper. It is illus-
trated with two different scenarios: a synthetic video
and a real one. It was decided not to choose a signifi-
cance level for the SKen test but to rank it by p-values
instead. This is due to the lack of precise information
on whether the amount of features selected as smooth
would be enough to run the 3D reconstruction algo-
rithm.
It is noteworthy that when a feature is not selected
by the SKen test, it does not necessarily mean that this
feature is not smooth; it just means it is less smooth
than the others selected. Likewise, if a feature is se-
lected, it does not necessarily mean that this feature
is perfectly smooth, but it is smoother than the other
unselected.
The system employed in this work for the 3D re-
construction is the R3D (Farias, 2012). This sys-
tem uses the 3D reconstruction pipeline described in
(Pollefeys, 1999), KLT tracker (Lucas and Kanade,
1981) and the feature detection executed with the
GFTT algorithm (Shi and Tomasi, 1994). Each recon-
struction algorithm has its peculiarities and the R3D,
for instance, needs at least thirty features tracked be-
tween the first and second keyframe. In other words,
thirty features must exist and be tracked in the first
two keyframes.
Applying a significance level α to SKen and se-
lecting the features considered inliers cannot guaran-
tee that the initial conditions of the R3D would be
satisfied. For this reason, it is necessary to provide as
input to the R3D the smoother features that meet the
requirements of such reconstruction algorithm. The
SKen test applied to R3D is given by the algorithm as
follows 2.
Algorithm 2: Application of SKen to the R3D system.
1. Apply the SKen test in the tracked features.
2. Sort in ascending order, i.e. by smoothness.
3. Select the first thirty features.
while until there is not 30 features that were tracked
from first to second keyframe do Add to the set the
next feature in the order of smoothness.
end while
4. This result is the new input set for the R3D.
In some reconstruction cases, the 3D mesh may be
visually harmed, because when the SKen is applied it
results in a drastic reduction in the amount of features.
In most cases, the 3D mesh is discarded at first, since
the most important requirement for 3D reconstruction
is a well-estimated camera pose (Farias, 2012). When
a good pose estimation is obtained, it can be generated
a 3D mesh from dense reconstruction algorithms with
better properties than the previous one(Furukawa and
Ponce, 2010).
In order to evaluate the quality of a 3D reconstruc-
tion, the reprojection error frame by frame must be
checked (Hartley and Zisserman, 2004). The finest
reconstruction is the one with the lowest average re-
projection error frame by frame. This parameter can
be calculated either as the average reprojection error
of all features in all frames or as the average of the av-
SKen:AStatisticalTestforRemovingOutliersinOpticalFlow-A3DReconstructionCase
205
erages of the reprojection error frame by frame; both
results are similar (Bolfarine and Bussab, 2005).
4.1 Synthetic Video
The video called Home (Figure 2) is a 640x480 pix-
els resolution video and it has 102 frames over ten
seconds.
Figure 2: Synthetic video - Home.
The initial step of the methodology consists in
tracking the scene. The best 1024 features were pro-
vided (Figure 3a) according with the GFTT classifi-
cation. The next step is to decompose the input data
provided by the tracker in (X, Y) as shown in the sec-
tion 3.5. It is important to perceive that in Figure 4,
the paths the features made basically follow a hori-
zontal movement, therefore it could be applied the
SKen test only on the X axis; it would considerably
decrease the processing time. Finally, the last step is
to apply the SKen in the 1024 features initially pro-
vided by the tracker and to sort the obtained p-values
as it increases. The lower the p-value, the smoother is
the path.
Figure 3: (a) Frame 1: Synthetic video - Home (b) Frame
1(SKen): Synthetic video - Home
When applying the methodology, the scene is re-
constructed using only the first 120 features, which
represents an 88.3 % drop concerning all features ini-
tially tracked. In other words, the scene was recon-
structed using the 120 smoother features and within
those 120 features there are thirty features that were
tracked between the first and second keyframe. The
smoother features evidenced by the SKen test can be
seen in Figure 3b.
If visually compared, it can be noticed that
throughout the tracked video with 1024 features,
some features did not present a smooth movement, as
shown in Figure 4a. The white rectangles in the figure
emphasize the most relevant non-smooth paths.
When adopted only the 120 features given by the
SKen test, these non-smooth features disappeared, as
shown in Figure 4b. This is due to the fact that these
features had a less smooth behavior, i.e. in the test
they presented a p-value higher than the final 120 se-
lected. Another example of non-smoothness also oc-
curs in Figure 5a. These same features do not appear
in the video sequence when using the SKen (Figure
5b), they also were classified as less smooth than the
selected by the SKen test.
Figure 4: (a) Frame 11: Synthetic video - Home (b) Frame
11 (SKen):Synthetic video - Home.
Figure 5: (a) Frame 88: Synthetic video - Home (b) Frame
88 (SKen): Synthetic video - Home.
As one the results of the methodology is the reduc-
tion of the number of features and thereby reducing
the processing time, the 3D mesh may be less detailed
than the one with the points initially tracked. How-
ever, as mentioned earlier in this section, the most im-
portant requirements are both the reprojection errors
frame by frame and a well-estimated camera pose.
Nevertheless, even with a restricted number of fea-
tures, the 3D mesh was not impaired. The mesh and
3D camera path can be seen in Figure 6.
Using the initial 1024 features, the video Home
presented an average reprojection error frame by
frame of 0.2327 pixels. When executed the SKen test,
the reconstructed scene had an average reprojection
error of 0.2180 pixels, which means it decreased by
6.3%. The reprojection errors frame by frame can be
seen in Figure 7.
Once the average reprojection error of all features
and the average error between features used by the
VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications
206
Figure 6: (a) 3D Reconstruction: Synthetic video - Home
(b) 3D Reconstruction (SKen): Synthetic video - Home.
Figure 7: Synthetic video - Home: Reprojection error frame
by frame.
SKen were very close to each other (only 6.3 % differ-
ence), it was performed the Wilcoxon test (Hollander
and Wolfe, 1999) for comparing these means. This
test was selected because the generated reprojection
error frame by frame, in this case, does not satisfy
aspects of normal distribution (Hollander and Wolfe,
1999). With the Wilcoxon test it was obtained a p-
value of 0.3764, which indicates that there is no sig-
nificant difference between the mean errors, i.e. the
reconstructed scene, when using the SKen test, did
not affect the reprojection error.
It must be highlighted that the Home video is a
synthetic composition; therefore it presents a low-
noise sequence. The main gain in this scenario is due
to a significant restriction in the amount of features
used, thereby also reducing the processing time of the
RANSAC. Even with an 88% decrease in the features,
it was statistically maintained the same reprojection
error.
The test was executed in an Intel(R) Core(TM) i3-
2120 CPU with 3.3GHz. The average execution time
of the SKen in C was approximately 0.4 milliseconds.
That is, the increment in runtime in C is practically
imperceptible.
4.2 Real Video
The second experiment used a real video called
Pineapple, which has been recorded with a 960x544
pixels resolution and composed of 164 frames over
five seconds (Figure 8).
As well as in previous case, the initial step of the
Figure 8: Real Video - Pineapple.
method consists in tracking the scene. The best 2000
features were provided (Figure 9a) according to the
GFTT classification. The next step is to decompose
the input data given by the tracker (X, Y) as shown in
section 3.5.
It is important to notice that, in the Figure 10,
the paths the features made basically follow vertical
movement; therefore it could be applied the SKen test
only on the Y axis. Again, the last step is to apply
the SKen in the 2000 features initially provided by
the tracker and to sort the obtained p-values as it in-
creases.
Figure 9: (a)Frame 1: Real Video - Pineapple (b) Frame 1
(SKen): Real Video - Pineapple.
Figure 10: (a) Frame 12:Real Video - Pineapple (b) Frame
12 (SKen): Real Video - Pineapple.
According to the Algorithm 2, the scene might be
reconstructed using only the first 153 features, which
represents a 92.3% decrease regarding the total fea-
tures that were initially tracked. In other words, the
scene was reconstructed using the smoother 153 fea-
tures and within those 153 features there are thirty
features that were tracked between the first and sec-
ond keyframe. The smoother features evidenced by
the SKen test can be seen in Figure 9b.
The resulting 3D mesh with the points selected
by the SKen test has shown slightly different changes
even though it used only 7.6% of the initial features.
The mesh and the 3D camera path can be seen in Fig-
ure 11.
Using the initial 2000 features, the Pineapple
SKen:AStatisticalTestforRemovingOutliersinOpticalFlow-A3DReconstructionCase
207
Figure 11: (a) 3D Reconstruction: Real Video - Pineapple
(b) 3D Reconstruction (SKen): Real Video - Pineapple.
video presented an average reprojection error frame
by frame of 1.689 pixels. When used the SKen test,
the reconstructed scene had an average reprojection
error of 0.90 pixels, which means a significant 46.9%
drop. The reprojection errors frame by frame can be
seen in Figure 12.
Figure 12: Real Video - Pineapple: Reprojection error
frame by frame.
Although the difference between the reprojection
errors was considerably high, and shown in Figure 12
(mainly from the frame 70 onward), it was performed
the Wilcoxon test (Hollander and Wolfe, 1999) for
statistical confirmation of this difference. As well as
in the previous case, this test was adopted because
the reprojection error frame by frame produced in
this case does not meet the conditions of the normal
distribution (Hollander and Wolfe, 1999). With the
Wilcoxon test it was obtained a p-value smaller than
10
5
, which indicates that there is significant differ-
ence between the mean errors, i.e. the reconstructed
scene reduced the average error of reprojection with
the SKen test.
The same experiment setup was defined, as well
as the hardware specifications. When implemented in
C the average processing time was approximately 0.5
milliseconds. That is, the increment in execution time
in C is practically imperceptible.
Considering this last scenario, applying the SKen
caused a 92.3% drop in the number of features. It
practically did not change the pipeline total runtime
and decreased by 46.9% the reprojection error frame
by frame.
5 CONCLUSIONS
This work presented important results concerning a
3D reconstruction pipeline by adopting a new tech-
nique called SKen. Although this method may reduce
the amount of features during the tracking stage, it is
not its objective; this is a consequence. The focus of
the SKen is the selection of the best features. Thereat,
the pipeline calculations can be more accurate and
thus more reliable scene reconstructions scenes will
be produced.
The two case studies present evidence that fea-
tures that performed smooth paths are inlier features.
This correlation can be confirmed by analyzing the
reprojection error behavior; with the SKen selected
features, reprojection errors either decreased or re-
mained the same. Furthermore, it is also shown that it
was not necessary to use the total amount of features
initially tracked, since the same reconstruction results
were achieved with less features and minor errors.
While in the synthetic and less noisy case the fea-
tures selected by the SKen test achieved the same av-
erage reprojection error than the features originally
tracked, this result was obtained by using only 11.7%
of features provided by the tracker. For the real case
with higher noise, the average reprojection error de-
creased by 46.9% and this result was obtained using
only 7.7% of features provided by the tracker.
The main challenge of this investigation was to de-
velop a methodology for obtaining results in real-time
as well as it should not compromise the performance
of the 3D reconstruction pipeline. The implementa-
tion of the SKen in C obtained an approximate av-
erage execution time of 0.5 milliseconds for a scene
with 2000 features. It represents only 1.5% of what
is necessary for meeting real-time requirements (ap-
proximately 33.3 milliseconds).
Finally, one of the most important features of the
proposed methodology is that the user does not need
to enter any parameters. Thus, this paper presented
an automatic and deterministic method in a way that
it does not need user intervention in order to remove
outliers.
REFERENCES
Barbosa, R. L. (2006). Caminhamento Fotogramtrico Uti-
lizando o Fluxo ptico Filtrado. PhD thesis, UNESP.
Bolfarine, H. and Bussab, W. O. (2005). Elementos de
Amostragem. Editora Edgard Blucher, So Paulo.
Bouguet, J. Y. (2000). Pyramidal implementation of the
lucas kanade feature tracker: Description of the algo-
rithm. Technical report, Intel Corporation.
VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications
208
Bussab, W. and Morettin, P. (2002). Estatstica Bsica. Edi-
tora Saraiva.
Choi, J. and Medioni, G. (2009). Starsac: Stable random
sample consensus for parameter estimation. IEEE In-
ternational Conference on Computer Vision and Pat-
tern Recognition, pages 675–682.
Farias, T. S. M. C. (2012). Metodologia para Reconstruo
3D Baseada em Imagens. PhD thesis, UFPE.
Fischler, M. A. and Bolles, R. C. (1987). Random sample
consensus: A paradigm for model fitting with appli-
cations to image analysis and automated cartography.
Morgan Kaufmann Publishers Inc.
Furukawa, Y. and Ponce, J. (2010). Accurate, dense and ro-
bust multiview stereopsis. IEEE Transactions on Pat-
tern Analysis and Machine Intelligence, 32(8):1362–
1376.
Hartley, R. I. and Zisserman, A. (2004). Multiple View Ge-
ometry in Computer. 2nd ed. Cambridge University
Press, ISBN: 0521540518.
Hming, K. and Peters, G. (2010). The structure from motion
reconstruction pipeline: a survey with focus on short
image. Kybernetika.
Hollander, M. and Wolfe, D. A. (1999). Nonparametric
Statistical Methods, 2nd Edition. Wiley-Interscience,
2 edition.
James, B. R. (2002). Probabilidade: Um curso em nvel
intermedirio. IMPA, Rio de Janeiro.
Lucas, B. D. and Kanade, T. (1981). An iterative image
registration technique with an application to stereo
vision (ijcai). In Proceedings of the 7th Interna-
tional Joint Conference on Artificial Intelligence (IJ-
CAI ’81), pages 674–679.
Nist
´
er, D. (2003). Preemptive ransac for live structure and
motion estimation. In Proceedings of the Ninth IEEE
International Conference on Computer Vision - Vol-
ume 2, ICCV ’03, pages 199–, Washington, DC, USA.
IEEE Computer Society.
Pollefeys, M. (1999). Self-calibration and metric 3D recon-
struction uncalibrated image sequeces. PhD thesis,
ESAT-PSI.
Rodehorst, V. and Hellwich, O. (2006). Genetic algorithm
sample consensus (gasac) - a parallel strategy for ro-
bust parameter estimation. In Computer Vision and
Pattern Recognition Workshop, 2006. CVPRW ’06.
Conference on, page 103.
Rousseeuw, P. and Leroy, A. (1987). Robust Regression and
Outlier Detection. John Wiley.
Shi, J. and Tomasi, C. (1994). Good features to track. In
1994 IEEE Conference on Computer Vision and Pat-
tern Recognition (CVPR’94), pages 593 – 600.
Torr, P. H. S. and Zisserman, A. (2000). Mlesac: A new
robust estimator with application to estimating image
geometry. Computer Vision and Image Understand-
ing, 78:2000.
SKen:AStatisticalTestforRemovingOutliersinOpticalFlow-A3DReconstructionCase
209