Stabilization of Endoscopic Videos using Camera Path from Global

Motion Vectors

Navya Amin, Thomas Gross, Marvin C. Ofﬁah, Susanne Rosenthal, Nail El-Sourani

and Markus Borschbach

Competence Center Optimized Systems, University of Applied Sciences (FHDW),

Hauptstr. 2, 51465 Bergisch Gladbach, Germany

Keywords:

Medical Image Processing, Endoscopy, Global Motion Vectors, Local Motion Estimation, Image Segmenta-

tion, Video Stabilization.

Abstract:

Many algorithms for video stabilization have been proposed so far. However, not many digital video stabiliza-

tion procedures for endoscopic videos are discussed. Endoscopic videos contain immense shakes and distor-

tions as a result of some internal factors like body movements or secretion of body ﬂuids as well as external

factors like manual handling of endoscopic devices, introduction of surgical devices into the body, luminance

changes etc.. The feature detection and tracking approaches that successfully stabilize the non-endoscopic

videos might not give similar results for the endoscopic videos due to the presence of these distortions. Our

focus of research includes developing a stabilization algorithm for such videos. This paper focusses on a spe-

cial motion estimation method which uses global motion vectors for tracking applied to different endoscopic

types (while taking into account the endoscopic region of interest). It presents a robust video processing

and stabilization technique that we have developed and the results of comparing it with the state-of-the-art

video stabilization tools. Also it discusses the problems speciﬁc to the endoscopic videos and the processing

techniques which were necessary for such videos unlike the real-world videos.

1 INTRODUCTION

Invasive diagnostics and therapy makes it necessary

for the surgeon or scientist to insert and move the

endoscopic devices within the inner parts of the hu-

man body. Based on the organ to be diagnosed, the

endoscopic procedure is classiﬁed as Bronchoscopy

(Daniels, 2009), Laryngoscopy (Koltai and Nixon,

1989), Gastroscopy, Laparoscopy (Koltai and Nixon,

1972), Rhinoscopy (University of Tennessee Col-

lege of Veterinary Medicine, 2012) and Colonoscopy.

Visualization of stable videos by physicians is very

essential during surgery. But the videos generated

during endoscopy are very distorted and shaky due

to the camera shakes occurring during insertion of

an endoscopic apparatus into the human body, the in-

ternal body movements (Breedveld, 2005) like heart-

beat, expansion and contraction of the organs as a

response to stimuli as well as secretion of body ﬂu-

ids like saliva, mucus, blood etc. Miniaturization

of the endoscopic devices also adds to the distor-

tions. This makes it difﬁcult for the surgeon to di-

agnose or operate. Hence, processing and stabiliza-

tion of such videos becomes necessary to overcome

such problems. Many mechanical video stabiliza-

tion systems are available for endoscopic video stabi-

lization commercially. These physical means of sta-

bilisation use either a gyroscope or some other self

stabilizing hardware device which holds the camera

(Chatenever et al., 2000). But the drawback is that

they are large, consume a lot of energy and are ex-

pensive. Another stabilzation system developed by

Canon also stabilizes before the video is converted

into digital data requiring larger and heavier cam-

era making it challenging for endoscopic cameras

(Canon, 1995). Digital video stabilization system is

a better, cheaper and compact alternative for the en-

doscopic purposes. Among the available state-of-the-

art digital video stabilization tools Adobe’s video sta-

bilization algorithm, which was subject to continu-

ous development in recent years, is one of the cur-

rent state-of-the-art stabilization algorithms. Adobe’s

algorithm is based on the well-known Kanade-Lucas-

Tomasi feature tracking (KLT) method, which detects

features in a start frame and tracks them over sev-

eral successive frames. The tracked feature point co-

130

Amin N., Gross T., C. Ofﬁah M., Rosenthal S., El-Sourani N. and Borschbach M..

Stabilization of Endoscopic Videos using Camera Path from Global Motion Vectors.

DOI: 10.5220/0004688801300137

In Proceedings of the 9th International Conference on Computer Vision Theory and Applications (VISAPP-2014), pages 130-137

ISBN: 978-989-758-003-1

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

ordinates form the trajectory paths, which represent

the movements of a video camera. In the subsequent

step, these shake-prone trajectories are smoothed with

a smoothing method to get a shake-free video. The

algorithm includes additional steps that are described

in detail in the recent publications (Liu et al., 2011),

(Liu et al., 2009) and (Zhou et al., 2013). Another

similar method is illustrated in the algorithm which

is integrated into YouTube’s video editor and is also

based on the KLT tracking method (Grundmann et al.,

2011). The method is explained in detail in (Grund-

mann et al., 2012). Microsoft’s algorithm worked

in a slightly different way as follows (Matsushita

et al., 2005), (Matsushita et al., 2006). The motion

in the video is estimated by means of the Hierarchical

model-based block-motion algorithm. This approach

does not track the features across multiple frames to

get the trajectories, but the camera path is determined

from the movements from every consecutive frame.

These motion vectors are used to create a trajectory,

which approximate the original camera path and still

contains the unwanted camera shake. This path is

then smoothed to obtain a shake-free video. Detailed

descriptions are found in (Bergen et al., 1992).

Research has been carried out on how well cur-

rent state-of-the-art stabilization tools and algorithms

perform on endoscopic videos. Luo et al (Lu

o et al.,

2010) tested different feature detection algorithms on

bronchoscopic videos. Wang et al used their Adaptive

Scale Kernel Consensus Estimator to estimate better

motion model parameters (Wang et al., 2008). There

exist many video stabilization tools - commercial as

well as freeware. However, application of these tools

for stabilizing endoscopic videos proved to be ineffec-

tive: Related work still shows that even high-quality

stabilization tools like the Google’s video stabiliza-

tion algorithm by Grundmann et al. (Grundmann

et al., 2011) used in Youtube Video Editor and the

Adobe Systems algorithm by Liu et al. (Liu et al.,

2011) appeared to perform with moderate to unsatis-

factory results, although these are the best among a set

of state-of-the-art tools (Ofﬁah et al., 2012c), (Ofﬁah

et al., 2012a),(Ofﬁah et al., 2012b).

2 STATEMENT OF PROBLEMS

The problems with the stabilization of endoscopic

videos are peculiar to the real world videos. In addi-

tion to the removal of high-frequency camera shakes

from normal videos, it is important that uneven cam-

era panning is aligned so that the output video looks

smooth to the viewer. High-frequency camera shakes

are also undesirable in endoscopic videos, but due

to the area of activity of an endoscope in the hu-

man body, sprawling camera movements always exist.

Also, in contrast to the real world video, endoscopic

recordings include usually no long video sequences,

but short scenes with quick image content changes.

Even small oscillating movements change the image

content to a large extent, because the image objects

are very close to the lens. Due to these circumstances

current state-of-the-art video stabilization algorithms

are not ideal for stabilizing endoscopic videos. No

long trajectories for endoscopic videos are available.

An algorithm has to be able to compensate the high-

frequency camera shake. To ensure a surgeon receives

a stable video, the removal of high-frequency shake is

the primary task of an endoscopic video stabilization

system. Involuntary movements, like movements of

body parts should not be removed whether the camera

is steady or navigating. The stabilization algorithm

by Microsoft deals with the stabilization of videos,

which determines the camera movements from frame-

to-frame. This approach is better suited for medical

videos. Adobe’s video stabilization uses the Content

Preserve Warp Technique and approximates the cam-

era movements with the information of the tracked

feature points. However, this algorithm has signiﬁ-

cant difﬁculty in texture-less image regions. Only a

few traceable feature points are found to determine

an accurate approximation for the image distortion,

which can lead to a strong erroneous warp of the im-

age content. Erroneous deformation is completely in-

admissible in medical images and thus the algorithm

used by Adobe is not favourable for stabilizing endo-

scopic videos.

Fast movements occur very frequently in endo-

scopic videos. This is because the objects are lo-

cated close to the camera lens leading to a very fast

change of the image content as well as motion blur.

In such frames many feature points are lost and a va-

riety of trajectories break off abruptly for which the

tracker must be reinitialized. In comparison to the

test recordings of our laboratory, endoscopic videos

include much less image features, like edges or tex-

ture, resulting in a smaller number of detected feature

points. For this reason, long trajectories are normally

not detected in case of endoscopic videos. This is a

big disadvantage for an endoscopic stabilization al-

gorithm, in which a large number of long trajectories

for a clean stabilization is essentially required. The

tracker must therefore be re-initialized much more

frequently for a stabilized recording. Since the fea-

ture detection is the ﬁrst and crucial step of video

stabilization, a maximum number of possible feature

points is required for a better image transformation

procedure. Multi-frame feature tracking approach to

StabilizationofEndoscopicVideosusingCameraPathfromGlobalMotionVectors

131

obtain long trajectories is used in the Adobe stabiliza-

tion algorithm. The long trajectories that such an al-

gorithm demands are not available in case of endo-

scopic videos. Thus, multi-frame feature tracking ap-

proach is not suitable endoscopic videos. Frame-to-

frame feature tracking as used in YouTube stabilizer

is the solution for this problem which has been used

in the current EndoStabf2f algorithm.

3 METHODS FOR VIDEO

STABILIZATION

The limitations of the previously proposed methods

for video stabilization and the designated problems

make it necessary to develop a new algorithm suited

for endoscopic videos. The videos are processed to

get rid of the meta-data, like text information in the

video frame and to determine the shape of the image

content if it has a circular, a polygon or a full frame

format (no black areas in the video frames). If the

video contains a circular output format or unwanted

text areas, then these regions are segmented and re-

moved prior to stabilization. These steps fall into the

area of image processing and signiﬁcantly contribute

to a better video stabilization.

3.1 Image Preprocessing

Because of different endoscopic types, the quality and

output format of the video recordings are signiﬁcantly

different. The endoscopic content in the videos can

be circular, rectangular or polygonal (Gershman and

Thomson, 2011). Furthermore, endoscopic videos

can not only contain image information but also ar-

eas with text. These are located either outside, dis-

played on the edge or within the video frame (Fig-

ure 2). The preprocessing step aims at achieving an

improved quality and information content of medical

images. This does not necessarily mean a visual im-

provement of the image material but an improvement

of the results with respect to subsequent processing

steps. For example, preprocessing enhances the qual-

ity of the videos to enable an improved and more ro-

bust feature detection in the frame. The following de-

scribes preprocessing steps which are included in our

video stabilization algorithm.

3.1.1 Segmentation of Undesirable Image

Regions

Some endoscopic videos contain meta-data like time

or year of capture or type of endoscopy inside the

video content. Text that appears in the video stays in

Original video

Preprocessing 1: Segmentation steps

Preprocessing 2: Image Enhancement

Video Stabilization Process

Frame 1 - n

1. ME: Feature Tracking &

Global Motion Estimation

2. Smoothing &

Motion Compensation

3. Image Composition

Stable Video

Optional: Additional Steps

Cut out black reagion

Figure 1: Video Stabilization workﬂow.

the same position throughout the video which could

be a disadvantage for a robust feature tracking. The

video in Figure 2 (Original Video Frame) contains

text information in the video frame. The strongest

features are detected only at this region in the video.

For this reason, a segmentation step is required to seg-

ment and remove the text information from the video

(See ﬁgure 2):

1. Create an averaged image from random video

frames. This serves as the input image for the bi-

nary mask. Many images are used randomly and

the image pixels of the video frame are averaged.

The areas containing endoscopic image informa-

tion are brighter. The dark black border around

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

132

the video image contains dark to black pixels and

hence is darker than the region containing endo-

scopic content. A better deﬁnition for creating the

binary mask is guaranteed.

2. Create a binary mask with chosen threshold.

3. Image processing steps: edge detection, dilation,

erode, segmentation and selection of the actual

endoscopic image content.

4. Reduce the size of the circular mask with x per-

cent to get a tolerance to the blurred edges of the

original video frame existing in some cases.

5. Optional: identiﬁcation of the maximum square

area within the video screen for better feature de-

tection.

Finally, one gets a video frame mask showing only

the important medical image content. Surrounding

regions are excluded. Another algorithm was im-

plemented, which automatically ﬁnds the maximum

square area within the video frame. (See Figure 2).

3.1.2 Removal of Grid Distortions from

Fiberglass-endoscopic-videos

Another automated preprocessing step is the detec-

tion of pixel grids in the video. Videos recorded with

the ﬁberglass endoscopes, contain a clearly visible

image grid due to the distinctive design of the en-

doscope, that covers the entire video frame. These

highly prominent pixels of the image grid are often

detected as robust feature points. Similarly, with the

text information, these grid pixels always remain in

the same location in the video frame. A robust fea-

ture tracking is thus signiﬁcantly affected. A prepro-

cessing step, using the Fourier transform, and addi-

tional image processing operations, remove the grid

from the endoscopic video.

3.2 Feature-based Motion Estimation

(ME Unit)

The movements of the endoscope are recorded us-

ing the KLT algorithm. But unlike the video stabi-

lization algorithm from Adobe and Google, this al-

gorithm does not use any anchor frames to track in-

dividual features over many frames to form feature

trajectories. Instead, the global and local motion of

the camera is determined only by successive frames,

using the KLT with our algorithm. Thus, the ME

unit determines the Global Motion Vectors (GMV)

between two adjacent frames of a video sequence.

From this frame-to-frame motion information a tra-

jectory is created, which represents the approximated

original camera path. This trajectory still contains the

undesirable camera shake. To compensate the shake,

a smoothing function is used to remove the jitters and

the irregular camera movement. The great advantage

of this method is that only two frames are needed to

determine the motion from one frame to another. For

example, in Adobe’s algorithm, 50 or more frames

are used to create a trajectory (Liu et al., 2011). As

seen in our experiments, such long trajectories do nor-

mally not appear in endoscopic videos. The crashing

of multi-frame feature tracking algorithm is described

in (Gross et al., 2014). The individual steps are de-

scribed in detail:

3.2.1 Global Motion Estimation

After the preprocessing is completed, the video stabi-

lization process begins. The ﬁrst frame of the video is

read. This serves as the initialization frame for fea-

ture detection. The feature points are detected us-

ing the KLT method within the previously determined

video mask (ROI). The output represents the x/y co-

ordinates of the feature points (FPs) found. Next, the

FPs in the second frame are detected. The two sets of

feature points, each consisting of x and y coordinates

are compared, and only valid feature points (feature

point pairs) are used. Feature points which were not

found in the previous frame are discarded. The global

motion in the image is then calculated by subtracting

the feature coordinate values of the ﬁrst frame from

the coordinate values of the second frame (only for

valid feature points). In each case, all detected FPs

coordinate values (separately for x and y) and motion

vectors are stored in a matrix. In the next step, the

second frame is used as the initialization frame and in

the third frame the FPs are tracked. These steps were

repeated until the entire video is processed. Always,

one frame serves as the initialization frame and the

adjacent frame serves as the tracking frame for mo-

tion estimation.

3.2.2 Local Motion Estimation

Sometimes local movements may appear in the video

which have nothing to do with the global camera mo-

tion. For example, such local movements are the re-

sult of the movement of the surgical devices. This

would affect the global motion estimation. Therefore,

these motions were explored in the second step and

were excluded from the calculation of the global mo-

tion. From all feature coordinates, a average global

mean value is calculated. This value represents the

average pixel difference of the current feature point

position and its position in the previous frame. From

the previously calculated motion vectors for each FP,

the Mean Absolute Deviation (MAD) (Sachs, 1984)

StabilizationofEndoscopicVideosusingCameraPathfromGlobalMotionVectors

133

Figure 2: Segmentation workﬂow. Left: Original image frame. Second left: Video mask with unnecessary regions deleted.

Third left: Inscribed square region in the video mask. Right: ROI for feature tracking in the video.

is calculated. This indicates how much a FP moved

away from its previous location. This is calculated

separately for the x and y component of the motion

vectors. The values which differ in + / - the MAD

value from the mean of all x and y values are dis-

carded and not used for the subsequent calculation of

the global camera motion.

MAD = mean

− mean

)| (1)

3.2.3 Case Distinction

• If no FPs are found in the frames:

In some cases it may occur that no feature points

are detected, such as when individual frames con-

tain too few or texture-less regions. In this case,

global motion cannot be calculated and the global

motion of the frame is set to zero. To use the

original frame could be an unsatisfactory solu-

tion, since the path of the global movement is

interrupted and as a result, the frames begin to

jump. Hence, the motion vectors from the pre-

vious frame are used to avoid the jump.

• Creation of the camera path:

The determined and stored global motion vectors

are used to create the camera path which repre-

sents the movement in the video (see ﬁgure 3).

As seen, the curve is still very irregular. The

ﬁne irregularities of the curve represent the high-

frequency camera shake in the video, which has

to be removed.

3.3 Smoothing

In the ME step all global motion vectors of the frames

are stored, and formed into a global camera path.

The task of the smoothing part is to determine a

smoother camera path curve to ensure a shake-free

video sequence. For smoothing the camera path,

a non-parametric kernel smoothing method with the

Nadaraya-Watson estimator is used to estimate an un-

known regression function (Cai, 2001). In ﬁgure 3,

0 50 100 150 200 250

−50

100

150

200

250

300

350

Figure 3: Motion trajectory extracted from a video. The

dashed line represents the shaky camera path, the smoothed

line the unshaky camera path.

the path of the global motion is shown in comparison

to the determined smoothed motion.

3.4 Motion Compensation and Image

Composition

To obtain a stabilized video, unwanted movements

were compensated by shifting the frames from the de-

tected feature point coordinate values obtained during

motion estimation to the the values determined in the

smoothing step (Figure 4). Further the existing black

regions as shown on ﬁgure 4 which are a result of the

frame-shifts are further removed to enable better and

clean visualization of the stable video. The black re-

gions are cropped out in case of rectangular videos.

In case of circular endoscopic videos, instead of the

extra black regions there exists a problem of jump-

ing circles as a result of the motion compensation. A

stabilized endoscopic video with jumping endoscopic

content is not good for visualization. Such videos

are further masked by calculating an average mask

across the whole video. A largest circle excluding the

jumping regions is calculated and the whole video is

masked using this mask. Many post-processing tech-

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

134

Figure 4: Motion compensated video frame.

niques for video stabilization are applied as per re-

quirements. In case of our endoscopic videos, harsh

cropping out of the black regions cause tremendous

loss of important information. Hence, the use of this

step completely depends on the user and the require-

ment.

4 EXPERIMENTS

We compare our stabilization algorithm (En-

doStabf2f) with Youtube (YT) and Adobe (AE). No

other digital stabilizers speciﬁcally for endoscopic

videos are available. Thus we used the best perform-

ing state-of-the-art digital stabilization tools for our

comparison experiments (Ofﬁah et al., 2012a). For

this, 11 test videos are used:

1. Two real world videos containing forward and

backward movement of the camera.

2. Nine endoscopic videos containing different types

of endoscopic distortions.

Videos are labelled according to their types (see Ap-

pendix). Videos are cropped into different sections to

obtain sub-videos containing forward and backward

movement, steady scenes, distortions like bubbles,

foreground moving objects and circular endoscopic

videos trimmed into rectangular regions. Videos are

stabilized using frame-to-frame motion estimation us-

ing KLT tracker and compensated using perspective

transformation.

The black regions in the edges which are a result of

the compensation are cropped out in case of rectangu-

lar videos. The size of the black region to be cropped

is decided on the basis of the maximum shift cal-

culated for the x and the y axis. This trimming is

done so as to make the resultant stabilized videos us-

ing our algorithm comparable to the other two algo-

rithms. Youtube scales the videos to get rid of the

black regions in the stabilized video. Adobe pro-

vides an option to apply this cropping. The exisitng

black regions would affect the calculated PSNR val-

ues. Since the presence of black regions would af-

fect the benchmarking results, these black regions are

cropped out in both Adobe and our stabilization algo-

rithm. The stabilized videos are compared across the

2 stabilization algorithms using Inter-frame Transfor-

mation Fidelity where Inter-Frame Peak-Signal-To-

Noise-Ratio (PSNR) is calculted for every stabilized

video.

5 RESULTS AND DISCUSSION

The results of stabilization vary for different types of

videos used. The Endostabf2f algorithm is designed

speciﬁcally for endoscopic videos unlike YT and AD.

The stabilization by the YT stabilizer results in com-

pressed video resulting in loss of quality, automatic

region of interest selection and scaling leading to im-

mense probable loss of some important medical in-

formation. Also, no customization of the stabilization

procedure is possible as per user’s requirements. Un-

compressed, better quality video without any loss of

information is the prime necessity of medical endo-

scopic videos. Thus, YT would not be successful for

stabilizing endoscopic videos as per the requirements

of the physician. Endostabf2f successfully stabilizes

the endoscopic videos fulﬁlling the above mentioned

requirements. We further compared our Endostabf2f

with Adobe AE where customization of the stabiliza-

tion procedure as per requirements is possible. How-

ever, the results show that the performance of the sta-

bilization procedure is better for EndoStabf2f (see ﬁg-

ure 5) when compared to Adobe AE stabilizer. The

videos used for benchmarking are preprocessed for

both the algorithms to enable the assessment of mo-

tion estimation and compensation. This results in bet-

ter PSNR values for AE. If the PSNR values are com-

pared without preprocessing, there is a possibility for

the values to be much lower due to the existing dis-

tortions like grid, moving circle etc. The PSNR val-

ues for the grid-removed bronchoscopic videos (g,h,l

and j) show that immense body movements in the

video affect the quality of video stabilization mak-

ing it worse. This is not in case of EndoStabf2f.

The real world video “Lab video 1” (f) which con-

tains forward and backward movements, is not well

stabilized by the AE algorithm resulting in a lower

PSNR value. On visualization, the shaky camera mo-

tion is not stabilized for a few frames. Similarly, in

case of other bronchoscopic and real world videos

(a,b,c,d and k) Endostabf2f gave better PSNR values

than AE. However, in case of the rhinoscopic video

with a steady camera (e) AE performed a bit better

StabilizationofEndoscopicVideosusingCameraPathfromGlobalMotionVectors

135

a b c d

I j k

Videos

PSNR values (in dB)

15 20 25 30 35

f2f AE

Figure 5: PSNR results for 11 videos.

than the Endostabf2f by 0.4 decibels approximately.

This might be because of the available long trajectory

for smoothening due to steady camera which usually

not frequently available in case of endoscopic videos.

Subjective assessment of the stabilized videos is re-

quired to afﬁrm the quality of stabilized videos as re-

quired by the user. Videos stabilized by AE contain

jumps in the frame during scene change. This is ex-

treme in case of endoscopic videos since they con-

tain very frequent scene changes. Hence, visualiza-

tion of such videos misguides the surgeon prooving

to be dangerous. Our algorithm in contrast takes these

issues into account and is customized for the purpose

of endoscopic video stabilization.

There is scope for further optimization of the En-

doStabf2f algorithm with respect to motion estima-

tion using optimized algorithms which is a part of our

ongoing research. In addition, the image composition

part could be optimized to exclude the black regions

after compensation without losing too much of impor-

tant information.

ACKNOWLEDGEMENTS

This work as a part of the PENDOVISION-project is

funded by the German Federal Ministry of Education

and Research (BMBF) under the registration identi-

ﬁcation 17PNT019. The ﬁnancial project organiza-

tion is directed by the Research Center J

ulich. The

work was conducted under supervision of Markus

Borschbach.

REFERENCES

Bergen, J. R., Anandan, P., Hanna, K. J., and Hingorani, R.

(1992). Hierarchical model-based motion estimation.

In ECCV ’92 Proc. of the Second European Confer-

ence on Computer Vision, pages 237–252. Springer-

Verlag London, UK.

Breedveld, P. (2005). A new, easily miniaturized steerable

endoscope. IEEE Engineering in Medicine and Biol-

ogy Magazine, Volume 24:40–47.

Cai, Z. (2001). Weighted nadaraya-watson regression esti-

mation. In Statistics and Probability Letters, volume

Volume 51, pages 307–318.

Canon (1995). Optical image Stabilizer. Canon U.S.A. Inc.,

http://www.usa.canon.com/cusa/consumer/standard-

display/Lens Advantage IS, Accessed:November 5,

2013.

Chatenever, D., Mattsson-Boze, D., and Amling, M. R.

(August 1, 2000). Image orientation for endoscopic

video displays. U.S. Patent 6097423 A.

Daniels, R. (2009). Delmar’s Guide to Laboratory and Di-

agnostic Tests. Delmar - Thomson Learning.

Gershman, G. and Thomson, M. (2011). Practical Pediatric

Gastrointestinal Endoscopy (1st Edition). Wiley, John

& Sons, Incorporated.

Gross, T., Amin, N., Ofﬁah, M. C., Rosenthal, S., El-

Sourani, N., and Borschbach, M. (2014). Optimiza-

tion of the endoscopic video stabilization algorithm

by local motion exclusion. In 9th International Joint

Conference on Computer Vision, Imaging and Com-

puter Graphics Theory and Applications, to appear.

Grundmann, M., Kwatra, V., Castro, D., and Essa, I. (2012).

Calibration-free rolling shutter removal. In Interna-

tional Conference on Computational Photography.

Grundmann, M., Kwatra, V., and Essa, I. (2011). Auto-

directed video stabilization with robust l1 optimal

camera paths. In IEEE Conference on Computer Vi-

sion and Pattern Recognition (CVPR).

Koltai, P. J. and Nixon, R. (1972). Laparoscopy-new instru-

ments for suturing and ligation. Fertility and Sterility,

Volume 23:274–277.

Koltai, P. J. and Nixon, R. (1989). The story of the laryn-

goscope. Ear, Nose, & Throat Journal 68, Volume

7:494–502.

Liu, F., Gleicher, M., Jin, H., and Agarwala, A. (2009).

Plane-based content-preservingwarps for video stabi-

lization. In SIGGRAPH 2009. ACM New York, NY,

USA.

Liu, F., Gleicher, M., Wang, J., Jin, H., and Agarwala, A.

(2011). Subspace video stabilization. ACM Transca-

tions on Graphics, Volume 30:1–10.

o, X., Feuerstein, M., Reichl, T., Kitasaka, T., and

Mori, K. (2010). An application driven comparison of

several feature extraction algorithms in bronchoscope

tracking during navigated bronchoscopy. In Proc. of

the 5th international conference on Medical imaging

and augmented reality, pages 475–484.

Matsushita, Y., Ofek, E., Tang, X., and Shum, H.-Y. (2005).

Full-frame video stabilization. In In Proc. Computer

Vision and Pattern Recognition, pages 50–57.

Matsushita, Y., Ofek, E., Tang, X., and Shum, H.-Y. (2006).

Full-frame video stabilization with motion inpainting.

Volume 28:1150–1163.

Ofﬁah, M. C., Amin, N., Gross, T., El-Sourani, N., and

Borschbach, M. (2012a). An approach towards a full-

reference-based benchmarking for quality-optimized

endoscopic video stabilization systems. In ICVGIP

2012, page 65.

Ofﬁah, M. C., Amin, N., Gross, T., El-Sourani, N., and

Borschbach, M. (2012b). On the ability of state-of-

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

136

the-art tools to stabilize medical endoscopic video se-

quences. In MedImage.

Ofﬁah, M. C., Amin, N., Gross, T., El-Sourani, N., and

Borschbach, M. (2012c). Towards a benchmarking

framework for quality-optimized endoscopic video

stabilization. In ELMAR, Sep.,2012 Proceedings,

pages 23–26.

Sachs, L. (1984). Applied Statistics: A Handbook of Tech-

niques. New York: Springer-Verlag.

University of Tennessee College of Veterinary Medicine, R.

(2012). http://www.vet.utk.edu/clinical/sacs/. inter-

nal/mips/rhinoscopy.php.

Wang, H., Mirota, D., Ishii, M., and Hager, G. D. (2008).

Robust motion estimation and structure recovery from

endoscopic image sequences with an adaptive scale

kernel consensus estimator. In CVPR.

Zhou, Z., Jin, H., and Ma, Y. (2013). Content-preserving

warps for 3d video stabilization. In IEEE Conference

on Computer Vision and Pattern Recognition.

APPENDIX

Table 1: List of the videos used for stabilization.

Video Description

a Bronchoscopic staboptic video of a rat with circular content

b Bronchoscopic staboptic video of a rat with rectangular content and moving camera

c Shaky video of a hippo

d Human Rhinoscopic 2 with rectangular content and steady camera

e Human Rhinoscopic 3 with rectangular content and steady camera

f Lab video 1 with forward and backward movement

g Bronchoscopic grid removed ﬁbreoptic video of a rat with steady camera

h Bronchoscopic grid removed ﬁbreoptic video of a rat with moving camera and distortion (Bubbles)

l Bronchoscopic grid removed ﬁbreoptic video of a rat with forward-backward movement of camera

j Bronchoscopic grid removed ﬁbreoptic video of a rat with rectangular content and steady camera

k Shaky video of a tiger with jittery motion

StabilizationofEndoscopicVideosusingCameraPathfromGlobalMotionVectors

137