NOVEL VIEW TELEPRESENCE WITH HIGH-SCALABILITY

USING MULTI-CASTED OMNI-DIRECTIONAL VIDEOS

Tomoya Ishikawa, Kazumasa Yamazawa and Naokazu Yokoya

Graduate School of Information Science, Nara Institute of Science and Technology, Ikoma, Nara, Japan

Keywords:

Telepresence, Novel view generation, Multi-cast, Network, Image-based rendering.

Abstract:

The advent of high-speed network and high performance PCs has prompted research on networked telepres-

ence, which allows a user to see virtualized real scenes in remote places. View-dependent representation,

which provides a user with arbitrary view images using an HMD or an immersive display, is especially effec-

tive in creating a rich telepresence. The goal of our work is to realize a networked novel view telepresence

system which enables multiple users to control the viewpoint and view-direction independently by virtual-

izing real dynamic environments. In this paper, we describe a novel view generation method from multiple

omni-directional images captured at different positions. We mainly describe our prototype system with high-

scalability which enables multiple users to use the system simultaneously and some experiments with the

system. The novel view telepresence system constructs a virtualized environment from real live videos. The

live videos are transferred to multiple users by using multi-cast protocol without increasing network trafﬁc.

The system synthesizes a view image for each user with a varying viewpoint and view-direction measured by

a magnetic sensor attached to an HMD and presents the generated view on the HMD. Our system can generate

the user’s view image in real-time by giving correspondences among omni-directional images and estimating

camera intrinsic and extrinsic parameters in advance.

1 INTRODUCTION

With high-speed network and high performance PCs,

networked telepresence, which allows a user to see

a virtualized real scene in a remote place, has been

investigated. Telepresence constructs a virtual envi-

ronment from real images, and therefore the user can

be given the feeling of richer immersion than using

a virtual environment constructed of 3D CG objects.

The technology can be widely applied to many en-

terprise systems for not only entertainments such as

virtual tours and games but also medical equipments,

educations, and so on.

In the past, we have proposed immersive telepres-

ence systems using an omni-directional video which

enables a user to control the view-direction freely

(Onoe et al., 1998; Ishikawa et al., 2005a). These

systems use a catadioptric omni-directional camera

or an omni-directional multi-camera system which

can capture omni-directional scenes and show view-

dependent images rendered by cutting out a part of

view-direction from an omni-directional image to the

users. Furthermore, one of the systems can show a

live remote scene by transferring an omni-directional

video via internet (Ishikawa et al., 2005a). We be-

lieve that view-dependent presentation is effective in

creating a rich presence and it will be a promising net-

worked media in the near future. However, presenta-

tion of novel views generated apart from the camera

positions is still a difﬁcult problem.

In this study, we aim at realizing a networked real-

time telepresence system which gives multiple users

synthesized live images by virtualizing a real scene

at an arbitrary viewpoint and in an arbitrary view-

direction. Especially, for providing users with a sense

of immersion, we virtualize the whole shooting scene

which includes both inside-out and outside-in obser-

vations from camera positions simultaneously. For

achieving high-scalability about increasing users, we

transfer live videos by using multi-cast protocol, so

148

Ishikawa T., Yamazawa K. and Yokoya N. (2007).

NOVEL VIEW TELEPRESENCE WITH HIGH-SCALABILITY USING MULTI-CASTED OMNI-DIRECTIONAL VIDEOS.

In Proceedings of the Ninth International Conference on Enterprise Information Systems - HCI, pages 148-156

DOI: 10.5220/0002359201480156

 SciTePress

that the system can support multiple users without in-

creasing network trafﬁc. By using this system, many

enterprise systems or applications would be devel-

oped such as a virtual tour system, which provides the

users with the feeling of walking in a remote sightsee-

ing spot, and a video conference system. We ﬁrst re-

view related works in the next section assuming such

applications, and then describe the detail of our sys-

tem.

2 RELATED WORK

One of the most famous telepresence systems, Quick-

TimeVR (Chen, 1995), gives a user an image in an

arbitrary view-direction generated by an image-based

rendering (IBR) technique from a panoramic image.

In this system, a static panoramic image is stitched

using a set of images acquired by a rotating camera.

Recently, telepresence systems using omni-

directional video cameras have been proposed. Uyt-

tendaele et al. (Uyttendaele et al., 2003) have pro-

posed a system which enables a user to change the

viewpoint and view-direction interactively like walk-

through in a real environment. It uses an omni-

directional multi-camera system for omni-directional

image acquisition and the user can control the view-

point using a joypad. Ishikawa et al. (Ishikawa

et al., 2005a) have developed immersive and net-

worked telepresence systems. The users can see

the video in an arbitrary view-direction depending

on their head movements by using an HMD attach-

ing a gyro sensor. In the above systems, the user’s

viewpoint is restricted at the position of the omni-

directional camera or onto the camera path.

For changing a viewpoint freely, a number of

methods for generating novel views from images ac-

quired by multiple cameras have been proposed. The

methods enable telepresence which provides a user

with a virtualized real scene in a remote site to give

a feeling of walking in a remote scene. Kanade et

al. (Kanade et al., 1997) and Saito et al. (Saito

et al., 2003) have proposed a method which generates

a novel view image in the 3D room. The 3D room

mounts a large number of cameras on the walls and

ceiling. These cameras can capture details of objects

but their methods need the assumption that the posi-

tions of objects are limited within the area covered

by the cameras. Therefore, these are suitable for 3D

modeling, rather than telepresence.

Koyama et al. (Koyama et al., 2003) have pro-

posed a method for novel view generation in a soccer

scene and their system has transmitted 3D video data

via internet. This method is based on the assump-

tion that the viewpoint is far from soccer players. The

method approximately generates a novel view image

using a billboarding technique in an online process. It

is difﬁcult to apply this method to the scenes where

dynamic objects are located close to cameras. In an-

other approach, a novel view generation method us-

ing a database of recorded parameterized rays pass-

ing through a space is proposed by Aliaga and Carl-

bom (Aliaga and Carlbom, 2001). This method ren-

ders a view image by resampling the rays through

an image plane from the database. It uses an omni-

directional camera for recording the rays efﬁciently

and the process is automated. This approach needs

dense rays acquired at different positions in an en-

vironment. It is difﬁcult to apply the method to dy-

namic environments. Zitnick et al.’s method (Zitnick

et al., 2004) generates high-quality view-interpolated

videos in dynamic scenes using a layered represen-

tation. This method estimates per-pixel depth by

segment-based stereo for each input image and al-

pha mattes are used for boundary smoothing. The

depth and color layer and the alpha matte layer are

composed by GPU functions. This method needs

high computational costs for depth estimation before

rendering. In previous work, we have proposed a

method which generates an omni-directional image

at a novel viewpoint from multiple omni-directional

images captured in a dynamic scene (Ishikawa et al.,

2005b). This method can support live video process-

ing and work in real-time. Owing to capturing images

with a wide ﬁeld of view by using multiple omni-

directional cameras, a novel viewpoint and a view-

direction are not restricted relatively compared with

the other methods. These features are suitable for

telepresence systems and we employ this method for

the present novel view telepresence system.

In this paper, we realize a networked telepresence

system which can provide multiple users with a sense

of immersion in a dynamic environment from live

video streams and has high-scalability taking advan-

tage of a feature of our novel view generation method.

Our system synthesizes views at users’ viewpoint and

in view-direction measured by a magnetic sensor at-

tached to an HMD and displays the generated view

images on the HMD in real-time. In experiments, we

demonstrate the feasibility of the system.

The rest of this paper is organized as follows.

Section 3 brieﬂy explains our novel view generation

method. Section 4 describes the prototype system of

novel view telepresence and the details. In Section

5, we show the experiment using the prototype sys-

tem and the results. Conclusions and future work are

ﬁnally given in Section 6.

NOVEL VIEW TELEPRESENCE WITH HIGH-SCALABILITY USING MULTI-CASTED OMNI-DIRECTIONAL

VIDEOS

149

Figure 1: Flowchart of novel view generation.

3 NOVEL VIEW GENERATION

In this section, we describe the novel view genera-

tion method from multiple omni-directional videos

(see Figure 1). This method segment input images

into static and dynamic regions and feasible techniqes

are applied to each region. In addition, for reducing

computational cost, it infers positions of dynamic re-

gions from multiple silhouette images. Firstly, omni-

directional images are acquired by omni-directional

cameras (Yamazawa et al., 1993) located in an envi-

ronment for capturing a wide area. Secondly, the ac-

quired images are segmented into static and dynamic

regions by using the robust background subtraction

technique (Yamazawa and Yokoya, 2003). Thirdly, a

static novel view image is generated by a morphing-

based technique, and a dynamic novel view image is

generated by computing visual hulls. Finally, the two

generated images are synthesized as a complete novel

view image. In Sections 3.1 and 3.2, we explain the

outline of novel view generation for static and dy-

namic regions, respectively.

3.1 Morphing-Based Rendering for

Static Regions

A morphing-based rendering technique is used to gen-

erate an omni-directional novel view image of static

regions from omni-directional images. This technique

needs 2D correspondences among input images. Note

that the corresponding points are given in advance.

The process of image generation based on morphing

is as follows.

Step1. The 3D positions of corresponding points are

computed by omni-directional stereo.

Step2. The 3D points computed by Step1 are pro-

jected onto the novel view image plane.

Step3. Triangulated patches in the omni-directional

novel view image are generated based on the pro-

jected points using Delaunay’s triangulation.

Step4. The omni-directional novel view image is ren-

dered by transforming and blending the parts of

input omni-directional images which correspond

to the triangle patches.

In this method, the calculation for pixels within

triangles and the blending are executed by using

OpenGL functions with GPU. Therefore, we can gen-

erate a novel view image without high CPU cost. For

more details, see the reference (Tomite et al., 2002).

3.2 Visual Hull Based Rendering for

Dynamic Regions

The novel view image of dynamic regions is gen-

erated by computing visual hulls. Visual hull con-

structed by the intersection of silhouette cones from

several viewpoint images deﬁnes an approximate ge-

ometric representation of an object. In general, the

shape-from-silhouette technique (Laurentini, 1994;

Seitz and Dyer, 1997) is used for computing the vi-

sual hull. The method employs the voxel represen-

tation of a space. Therefore, the cost of computing

the visual hull and the size of data become huge for

a wide area. Our proposed method uses the Image-

based Visual Hull technique (Matusik et al., 2000) for

computing visual hulls with reduced cost. The tech-

nique does not use voxels but generates a novel view

image by estimating the penetration of visual hull by

the ray from the novel viewpoint. The overview of the

process is as follows.

Step1. For each pixel in the novel view image, a ray

connecting a pixel and the novel viewpoint is pro-

jected onto input omni-directional images.

Step2. On the line of the ray, we search for a segment

in which all of the projected lines intersect with

dynamic regions. If it is found, the ray is judged

to penetrate the visual hull. If it is not found, the

ray does not penetrate the visual hull.

Step3. If the ray penetrates the visual hull, the pixel

in novel view image is colored. The point on the

intersection segment nearest to the novel view-

point is projected onto the input image that is se-

lected by the similarity of the angle formed by the

novel viewpoint, the intersection point, and the

ICEIS 2007 - International Conference on Enterprise Information Systems

150

Figure 2: Conﬁguration and ﬂow diagram of prototype system.

viewpoint of input image. The color of the pro-

jected point is decided as the color of pixel in the

novel view.

Step4. The process consisting of Step1-3 is executed

for all the pixels in the novel view image.

When all the pixels in the novel view image are

determined by this process, the computational cost is

large. For reducing the cost, we compute the regions

in the novel view images to which dynamic objects

are projected. The projected regions are decided as

object regions in the environment which are calclated

from multiple silhouettes in the input images. See the

reference (Morita et al., 2003) for the details of object

position estimation. We need to estimate the visual

hull only on the projected regions.

4 PROTOTYPE TELEPRESENCE

SYSTEM

In this section, we describe a prototype of telepres-

ence system which enables multiple users to see a

virtualized remote scene at an arbitrary viewpoint

and in an arbitrary view-direction interactively. The

prototype system transfers multiple omni-directional

videos captured in a remote site via network and the

users see novel view images generated from the live

videos. Our view generation method can generate

view images from the same set of omni-directional

images independently of user’s viewpoints and view-

directions. The prototype system takes advantage

of this feature and realizes high-scalability. In the

concrete, this system tranfers multiple videos using

multi-cast protocol in a network. Because of multi-

cast protocol, video packets are cloned by router or

hub automatically. The system then generates view

images from the videos. In the following, the conﬁg-

uration of the system and the process ﬂows are pre-

sented.

4.1 Conﬁguration of System

Figure 2(top) illustrates a conﬁguration of the proto-

type system. This system consists of the live video

server, which transfers multiple omni-directional

videos captured in a remote site, and the novel view

generators, which receive the transferred videos and

generate novel views. The details of conﬁgurations in

remote and user sides are described below.

[Live video server]

We arrange three omni-directional cameras (Suekage

NOVEL VIEW TELEPRESENCE WITH HIGH-SCALABILITY USING MULTI-CASTED OMNI-DIRECTIONAL

VIDEOS

151

Figure 3: Magnetic sensor (left) and receiver attached HMD

(right).

Inc. SOIOS55-cam) with IEEE1394 connectors and

connect them to the live video server in the present

implementation. The server is connected with Gigabit

Ethernet (1000BASE-T).

[Novel view generator]

The novel view generator computes a novel view

image according to the user’s viewpoint and view-

direction measured by a magnetic sensor (Polhemus

Inc. 3SPACE FASTRAK). The magnetic sensor mea-

sures the receiver’s 6DOF (positions and orienta-

tions) data. For measuring user’s viewpoint and view-

direction widely, we employ LongRanger (See Figure

3(left)) as a transmitter. The user wears a receiver at-

tached HMD (Olympus Inc. FMD-700, See Figure

3(right)) and sees displayed view images like walk-

ing in a remote site over changing the viewpoint and

view-direction. The generator also supports keyboard

and mouse operations for changing the viewpoint and

view-direction and displaying the generated images

onto an LCD monitor.

Table 1 summarizes the details of components of

the prototype system.

4.2 Process Flow

In our telepresence system, the live video server and

the novel view generators generate pre-computable

data before online processing for efﬁciently carrying

out online processing. Furthermore, the server carries

out the background subtraction process which does

not depend on the user’s viewpoint. We can distribute

the computational cost and the cost is not increased

when the view generators are added in the system.

In the transmission of data between the server and

the generator, we use multi-cast protocol for realizing

high-scalablity. In the following, we describe ofﬂine

and online processes according to Figure 2(bottom).

Firstly, pre-computable data are generated in the

Table 1: The details of prototype system components.

Omni-directional camera Resolution:

640x480[pixel]

Max. Frame rate:

15[fps]

Field of view:

Horizontal:360[degree]

Vertical:62[degree]

Live video server CPU:

Intel Pentium4 3.2GHz

Novel view generator1 CPU:

Intel PentiumD 940

GPU:

nVidia GeForce7300GS

Novel view generator2 CPU:

Intel Pentium4 3.2GHz

GPU:

nVidia GeForce6600GT

ofﬂine process.

[Live video server]

• Multiple omni-directional cameras arranged in a

remote place should be calibrated because our

novel view generation method needs accurate po-

sition and posture data of cameras. We cali-

brate the focal length, positions and postures of

the camera by minimizing the re-projection er-

ror of markers in the environment. Our approach

is similar to the method proposed by Negishi et

al. (Negishi et al., 2004). Note that the 3D po-

sitions of markers are measured by using a laser

rangeﬁnder and the markers are placed in all az-

imuth around cameras for avoiding deviations of

estimated positions.

• Invalid region mask is deﬁned because an omni-

directional image includes invalid regions, which

represent a projected camera part and corners of

an image. All of processes ignore the masked re-

gion.

[Novel view generator]

• Correspondences among input omni-direcitonal

images are given manually for morphing-based

rendering (Section 3.1). We assume that most re-

gions in the environment are static for a long time.

• The visual hull based rendering requires ray vec-

tors from a novel viewpoint to each pixel in a

novel view image. The vectors are computed in

the ofﬂine process for reduction of computaional

cost.

After the ofﬂine process, the system generates

novel view images time by time in the online process.

ICEIS 2007 - International Conference on Enterprise Information Systems

152

[Live video server]

1. The server acquires multiple omni-directional im-

ages. We employ DirectShow functions provided

by Microsoft Inc. for image capture.

2. Dynamic regions in each omni-directional image

are extracted by adaptive background subtraction

(see (Yamazawa and Yokoya, 2003) for details).

The results are deﬁned as binary dynamic region

masks (1bit/pixel).

3. The acquired images are encoded into the JPEG

format. Then, the encoded images and the dy-

namic region masks are transferred to novel view

generators by using multi-cast protocol. The

packets of transferred data are cloned by routers

or hubs automatically.

[Novel view generator]

1. The novel view generators receive multiple en-

coded images and dynamic region masks and then

generate a novel view image after taking a user’s

viewpoint and view-direction from a magnetic

sensor. To reduce the computational cost of view

generation, we restrict the computation not only

to the region described in Section 3.2 but also to

the pixels of regions included in the user’s view-

direction.

2. After the novel view generation, the generated

omni-directional image is transformed into a com-

mon planar perspective image. In this process, the

generated image on the frame buffer of GPU is

used as a textureimage directly and the planar per-

spective transformation uses GPU functions for

high-speed processing.

3. The transformed image is displayed onto an

HMD.

5 EXPERIMENTS

We have generated novel view images from live

videos acquired by omni-directional cameras in an in-

door environment. In this experiment, two users see

the remote site using two novel view generators inde-

pendently. One of the users (user1) is presented view-

dependent images by a magnetic sensor and an HMD

and changes the viewpoint and the view-direction

freely by walking. The other user (user2) sees novel

view images on an LCD monitor and changes the

viewpoint and view-direction by keyboard and mouse

operations. The omni-directional cameras are posi-

tioned forming a triangle (see Figure 4(top)). The es-

timated camera positions and postures are shown in

Figure 4: Remote omni-directional cameras (top) and esti-

mated camera positions (bottom).

Figure 4(bottom). The black points and three pyra-

mids mean the calibration markers and the cameras,

respectively. The distance between cameras is about

2m. The 70 corresponding points are manually estab-

lished by mouse clicking only for the static scene.

Figures 5 and 6 show appearances of both remote

and user sites and generated omni-directional novel

view images as well as planar perspective images.

The remote site in Figure 5 does not include dynamic

objects. On the other hand, the scene of Figure 6 in-

cludes two dynamic objects.

The user can change the viewpoint and view-

direction freely and can sense the feeling of existing

in the remote site. The resolution of omni-directional

novel view is 512x512 pixels and the ﬁeld of view

of the generated omni-directional image is same as

input. The effective resolution of the planar perspec-

tive view is about 30,000 pixels in the experiment and

the ﬁeld of view is 60 degrees. Each frame of novel

views is generated in about 200[ms]; that is, the frame

rate is about 5[fps]. The computation cost consists

of 80 milliseconds for receiving images and masks

and JPEG decoding, 20 milliseconds for the static

novel view generation, and 100 milliseconds for the

dynamic novel view generation. However, the time

for dynamic novel view generation is not constant be-

cause the time depends on the region size of dynamic

objects. From Figure 5, we can conﬁrm that the novel

view images are successfully generated according to

NOVEL VIEW TELEPRESENCE WITH HIGH-SCALABILITY USING MULTI-CASTED OMNI-DIRECTIONAL

VIDEOS

153

(a) Appearance of remote

site

(b) Appearance of user’s

site

novel view

(d) Planar perspec-

tive view

Figure 5: Results of novel view telepresence and experimental environment without dynamic objects.

(a) Appearance of re-

mote site

(b) Appearance of

user’s site

novel view

(d) User1’s view

image

(e) User2’s view

image

Figure 6: Results of novel view telepresence and experimental environment with dynamic objects (one of two persons is out

of view in (a)).

ICEIS 2007 - International Conference on Enterprise Information Systems

154

Figure 7: Network trafﬁc since starting server program.

the user’s viewpoint. From Figure 6, we can conﬁrm

that each user can see the same remote site from the

different viewpoint simultaneously. The network traf-

ﬁc of transmission from server is shown in Figure 7

for the experiment of Figure 6. The live video server

was started to transfer images and masks at time 0

and then, 6 seconds later, user1 started telepresence.

User2 joined in the system at 30 seconds after start-

ing server. In spite of increasing the number of users,

the network trafﬁc is constant. A server which is not

connected by clients has to send data in the present

implementation. The server should stop sending data

if any clients do not connect to it.

For use of the multi-cast protocol, multiple videos

are transferred as packets asynchronously so that re-

ception control of them has become difﬁcult. A small

number of packets is better for the reception control

but each image has to be compressed by high ratio. It

means that multiple videos become low quality. The

easiness of reception control and the quality of videos

are trade-off. In this experiment, we decided the rela-

tion empirically.

6 CONCLUSIONS

We have proposed a novel view telepresence sys-

tem with high-scalability using multi-casted omni-

directional videos. In our system, multiple users can

see a virtualized remote site at an arbitrary view-

point and in an arbitrary view-direction simultane-

ously. Furthermore, the network trafﬁc is constant

when increasing the number of users. In experiments

with the prototype system, we have conﬁrmed that our

system can give the feeling of walking in a remote site

to the users. When the viewpoint is in the middle of

the cameras and is close to dynamic objects, the qual-

ity of novel view image becomes low. This is caused

by the geometric error in approximate representation

using visual hulls with few cameras. The future work

includes improvement of quality of novel view image

and development of a method for giving point corre-

spondences among input images automatically.

REFERENCES

Aliaga, D. G. and Carlbom, I. (2001). Plenoptic stitching:

A scalable method for reconstructing 3D interactive

walkthroughs. In Proc. of SIGGRAPH2001, pages

443–451.

Chen, S. E. (1995). Quicktime VR: An image-based ap-

proach to virtual environment navigation. In Proceed-

ings of SIGGRAPH’95, pages 29–38.

Ishikawa, T., Yamazawa, K., Sato, T., Ikeda, S., Nama-

mura, Y., Fujikawa, K., Sunahara, H., and Yokoya, N.

(2005a). Networked telepresence system using web

browsers and omni-directional video streams. In Proc.

SPIE Electronic Imaging, volume 5664, pages 380–

387.

Ishikawa, T., Yamazawa, K., and Yokoya, N. (2005b).

Novel view generation from multiple omni-directional

videos. In Proc. IEEE and ACM Int. Symp. on Mixed

and Augmented Reality, pages 166–169.

Kanade, T., Rander, P., and Narayanan, P. J. (1997). Vir-

tualized reality: Constructing virtual worlds from real

scenes. IEEE MultiMedia Magazine, 1(1):34–47.

Koyama, T., Kitahara, I., and Ohta, Y. (2003). Live mixed-

reality 3D video in soccer stadium. In Proc. of IEEE

and ACM Int. Symp. on Mixed and Augmented Reality,

pages 178–187.

Laurentini, A. (1994). The visual hull concept for

silhouette-based image understanding. IEEE Trans.

on Pattern Analysis and Machine Intelligence,

16(2):150–162.

Matusik, W., Buehler, C., Raskar, R., Gortler, S. J., and

McMillan, L. (2000). Image-based visual hulls. In

Proc. of SIGGRAPH2000, pages 369–374.

Morita, S., Yamazawa, K., and Yokoya, N. (2003).

Networked video surveillance using multiple omni-

directional cameras. In Proc. 2003 IEEE Int. Symp. on

Computational Intelligence in Robotics and Automa-

tion, pages 1245–1250.

Negishi, Y., Miura, J., and Shirai, Y. (2004). Calibration

of omnidirectional stereo for mobile robots. In Proc.

of 2004 IEEE/RSJ Int. Conf. on Intelligent Robots and

Systems, pages 2600–2605.

Onoe, Y., Yamazawa, K., Takemura, H., and Yokoya,

N. (1998). Telepresence by real-time view-

dependent image generation from omni-directional

video streams. Computer Vision and Image Under-

standing, 71(2):154–165.

Saito, H., Baba, S., and Kande, T. (2003). Appearance-

based virtual view generation from multicamera

videos captured in the 3-D room. IEEE Trans. on Mul-

timedia, 5(3):303–316.

NOVEL VIEW TELEPRESENCE WITH HIGH-SCALABILITY USING MULTI-CASTED OMNI-DIRECTIONAL

VIDEOS

155

Seitz, S. M. and Dyer, C. R. (1997). Photorealistic scene

reconstruction by voxel coloring. In Proc. of IEEE

Conf. on Computer Vision and Pattern Recognition,

pages 1067–1073.

Tomite, K., Yamazawa, K., and Yokoya, N. (2002). Ar-

bitrary viewpoint rendering from multiple omnidirec-

tional images for interactive walkthroughs. In Proc.

of 16th IAPR Int. Conf. on Pattern Recognition, vol-

ume 3, pages 987–990.

Uyttendaele, M., Criminisi, A., Kang, S. B., Winder,

S., Szeliski, R., and Hartley, R. (2003). Image-

based interactive exploration of real-world environ-

ments. IEEE Computer Graphics and Applications,

24(3):52–63.

Yamazawa, K., Yagi, Y., and Yachida, M. (1993). New

real-time omni-directional image sensor with hyper-

boloidal mirror. In Proc. 8th Scandinavian Conf. on

Image Analysis, volume 2, pages 1381–1387.

Yamazawa, K. and Yokoya, N. (2003). Detecting moving

objects from omni-directional dynamic images based

on adaptive background subtraction. In Proc. 10th

IEEE Int. Conf. on Image Processing, volume 3, pages

953–956.

Zitnick, C. L., Kang, S. B., Uyttendaele, M., Winder, S., and

Szeliski, R. (2004). High-quality video view interpo-

lation using a layered representation. ACM Trans. on

Graphics, 23(3):600–608.

ICEIS 2007 - International Conference on Enterprise Information Systems

156