﻿Spatio-temporal Reflectance Sharing for
Relightable 3D Video
Naveed Ahmed, Christian Theobalt, and Hans-Peter Seidel
MPI Informatik, Saarbrücken, Germany,
[nahmed,theobalt,hpseidel]@mpi-inf.mpg.de,
WWW home page: http://www.mpi-inf.mpg.de/
Abstract. In our previous work [21], we have shown that by means
of a model-based approach, relightable free-viewpoint videos of human
actors can be reconstructed from only a handful of multi-view video
streams recorded under calibrated illumination. To achieve this purpose,
we employ a marker-free motion capture approach to measure dynamic
human scene geometry. Reflectance samples for each surface point are
captured by exploiting the fact that, due to the person’s motion, each
surface location is, over time, exposed to the acquisition sensors under
varying orientations. Although this is the first setup of its kind to measure
surface reflectance from footage of arbitrary human performances,
our approach may lead to a biased sampling of surface reflectance since
each surface point is only seen under a limited number of half-vector
directions. We thus propose in this paper a novel algorithm that reduces
the bias in BRDF estimates of a single surface point by cleverly taking
into account reflectance samples from other surface locations made of
similar material. We demonstrate the improvements achieved with this
spatio-temporal reflectance sharing approach both visually and quantitatively.
1 Introduction
The capturing of relightable dynamic scene descriptions of real-world events requires
the proper solution to many different inverse problems. First, the dynamic
shape and motion of the objects in the scene have to be captured from multiview
video. Second, the dynamic reflectance properties of the visible surfaces
need to be estimated. Due to the inherent computational complexity, it has not
been possible yet to solve all these problems for general scenes. However, in previous
work [21] we have demonstrated that the commitment to an adaptable a
priori shape model enables us to reconstruct relightable 3D videos of one specific
type of scene, namely of human actors. By means of a marker-free optical motion
capture algorithm, it becomes possible to measure both the shape and the
motion of a person from multiple synchronized video streams [2]. If the video
footage has, in addition, been captured under calibrated lighting conditions, the
video frames showing the moving person not only represent texture samples, but
actually reflectance samples. Since a description of time-varying scene geometry
II
is at our disposition, we know under what different incoming light and outgoing
viewing directions each point on the body surface is seen while the person
is moving. It thus becomes feasible to fit to each point on the body surface a
static parametric BRDF to describe the material properties, and a time-varying
normal to describe the dynamic change in surface geometry. Although we have
shown in our previous work that it is feasible to reconstruct dynamic surface
reflectance properties using only eight cameras and a static set of light sources,
this type of sensor arrangement leads to a biased sampling of the reflectance
space. Due to the fixed relative arrangement of lights and cameras, each surface
point is only seen under a limited number of half-vector directions. Furthermore,
even if an actor performs very expressive motion, surface points will never be
seen under all possible relative orientations to the cameras. This bias in the
reflectance data leads to a bias in the measured reflectance models which may
lead to an unnatural appearance if the 3D videos are rendered under virtual
lighting conditions that are starkly different from the measurement setup. We
thus propose spatio-temporal reflectance sharing, a method to cleverly combine
dynamic reflectance samples from different surface points of similar material during
BRDF estimation of one specific surface point. The guiding idea behind the
approach is to exploit spatial coherence on the surface to obtain more samples
for each texel while not compromising the estimation of spatially-varying details
in surface appearance. Temporal coherence is also exploited, since samples are
collected by combining measurements from subsequent time steps.
We continue with a review of the most relevant related work in Sect. 2. The
acquisition setup, the model-based motion estimation approach, as well as the
basic principles of dynamic reflectance estimation are described in Sect. 3. In
Sect. 4, we describe the nuts and bolts of our proposed dynamic reflectance
sharing approach and show how it fits into the original pipeline. Finally, in
Sect. 5 we demonstrate both visually and quantitatively that our novel sampling
strategy leads to improved results. We conclude in Sect. 6 and give an outlook
to possible future work.
2 Related Work
There is a huge body of literature on the estimation of reflectance properties
of static scenes from images that are captured under calibrated setups of light
source and camera. Typically, parameters of a BRDF model are fitted to the
data [18, 11] or appearance under novel lighting conditions is created via interpolation
between the images themselves [15]. A combination of reflectance estimation
and shape-from-shading to refine the geometry of static scenes is also feasible
[25, 17, 1, 4, 5]. In an independent line of research, many methods to capture
and render 3D videos of real-world scenes have been developed in recent years.
A popular category of algorithms employs the shape-from-silhouette principle to
reconstruct dynamic scene geometry by means of voxels, polyhedrals or point
primitives [14, 24, 13, 6, 12]. By finding temporal correspondences between pertime-step
reconstructions, it becomes feasible to generate novel animations as
well [19]. Another category of approaches reconstructs dynamic scene geometry
by means of multi-view stereo [27, 9, 22]. In any case, time-varying textures for
rendering are assembled from the input video streams. In contrast, the authors
in their previous work have proposed a model-based approach to free-viewpoint
video of human actors that jointly employs a marker-free motion capture method
and a dynamic multi-view texture generation approach to produce novel viewpoint
renditions [2, 20]. Unfortunately, none of the aforementioned methods can
correctly reproduce 3D video appearance under novel simulated lighting conditions.
Only few papers have been published so far that aim at relighting of dynamic
scenes. In [8], a method to generate animatable and relightable face models from
images taken with a special light stage is described. Wenger et al. [23] extend the
light stage device such that it enables capturing of dynamic reflectance fields.
Their results are impressive, however it is not possible to change the viewpoint
in the scene. Einarsson et. al. [3] extends it further by using a large light stage,
a trade-mill where the person walks on, and light field rendering for display.
Eventually, human performances can be rendered from novel perspectives and
relit. Unfortunately the method can only present single periodic motion, such as
walking, and is only suitable for low frequency relighting.
In contrast, the authors have proposed a model-based method for reconstructing
relightable free-viewpoint videos that extends measurement principles
for static parametric reflectance models to dynamic scenes [21]. For our 3D video
scenario, we prefer a compact scene description based on parametric BRDFs that
can be reconstructed in a fairly simple acquisition facility. This paper proposes
a novel solution to one important subproblem in the overall process, namely the
clever sampling of surface reflectance in order to minimize the bias in the estimated
BRDFs. This work has been inspired by the reflectance sharing method
of Zickler et al. to reconstruct appearance of static scenes [26]. By regarding
reflectance estimation as a scattered interpolation problem, they can exploit spatial
coherence to obtain more reliable surface estimate. Our algorithm exploits
both spatial and temporal coherence to reliably estimate dynamic reflectance.
However, since a full-blown scattered data interpolation would be illusive with
our huge sets of samples, we propose a faster heuristic approach to reflectance
sharing.
3 Relightable Free-viewpoint Video of Human Actors -
Preliminaries
The algorithm presented in this paper is a methodical improvement of one important
step within a larger framework to reconstruct and render relightable
free-viewpoint videos of human actors [21]. Although the algorithmic details of
this framework, as a whole, are not the subject of this paper, for better understanding
in the following we briefly elaborate on the acquisition setup used, as
well as the employed model-based marker-less motion capture algorithm. Thereafter,
we describe the basic principles of dynamic reflectometry that was used in
III
IV
(a) (b) (c)
Fig. 1. (a) Input frame, (b) body model in same pose, and (c) silhouette matching.
the original pipeline in order to motivate where the novel algorithm described
in Sect. 4 comes into play.
3.1 Acquisition
Inputs to our method are synchronized multi-view video sequences captured with
eight calibrated cameras that feature 1004x1004 pixel image sensors and record
at 25 fps. The cameras are placed in an approximately circular arrangement
around the center of the scene which is illuminated by two calibrated spot lights.
Since we conceptually separate the BRDF estimation from the estimation of the
dynamic normal maps, we record two types of multi-view video sequence for
each person and each type of apparel. In the first type of sequence, the so-called
reflectance estimation sequence (RES), the person performs a simple rotation
if front of the acquisition setup. One RES is recorded for each actor and each
type of apparel, and it is later used to reconstruct the per-texel BRDF models.
In the second type of sequence, the so-called dynamic scene sequence (DSS),
the actor performs arbitrary movements. Several DSS are recorded, and from
each of them, one relightable free-viewpoint video clip is reconstructed. Also the
second component of our dynamic reflectance model, the dynamic normal maps,
are reconstructed from each DSS.
3.2 Reconstructing Dynamic Human Shape and Motion
We employ an analysis-through-synthesis approach to capture both shape and
motion of the actor from multi-view video footage without having to resort to
optical markers in the scene. It employs a template human body model consisting
of a kinematic skeleton and a single-skin triangle mesh surface geometry [2,
20]. In an initialization step, the shape and proportions of the template are
matched to the recorded silhouettes of the actor. After shape initialization, the
model is made to follow the motion of the actor over time by inferring optimal
Fig. 2. Steps to estimate per-texel BRDFs.
pose parameters at each time step of video using the same silhouette matching
principle, Fig. 1. We apply this dynamic shape reconstruction framework to every
time step of each captured sequence, i.e. both RES and DSS. This way, we
know for each time step of video the orientation of each surface point with respect
to the acquisition setup which is a precondition for the subsequent dynamic
reflectometry procedure.
3.3 Dynamic Reflectometry
Our dynamic reflectance model consists of two components, a static parametric
isotropic BRDF for each surface point [16, 10], as well as a description of the timevarying
direction of the normal at each surface location. The first component of
the reflectance model is reconstructed from the video frames of the reflectance
estimation sequence, the second component is reconstructed from each dynamic
scene sequence. We formulate BRDF reconstruction as an energy minimization
problem in the BRDF parameters [21]. This minimization problem has to be
solved for each surface point separately.
The energy functional measures the error between the recorded reflectance
samples of the point under consideration and the predicted surface appearance
according to the current BRDF parameters. Given estimates of the BRDF parameters,
we can also refine our knowledge about surface geometry by keeping
the reflectance parameters fixed and minimizing the same functional in the normal
direction, Fig 2. Once the BRDF parameters have been recovered from the
RES, a similar minimization procedure is used to reconstruct the time-varying
normal field from each DSS.
Before estimation commences, the surface model is parameterized over the
plane and all video frames are transformed into textures. The estimation process
is complicated by the fact that our shape model is only an approximation.
Furthermore, potential shifting of the apparel over the body surface while the
person is moving contradicts our assumption that we can statically assign material
properties to individual surface points. We counter the first problem by
means of an image-based warp correction step, and solve the latter problem by
detecting and compensating textile shift in the texture domain. For details on
each of these steps, please refer to [21].
In the original pipeline, as it was summarized above, we have estimated
BRDF parameters for each surface point by taking only reflectance samples of
V
VI
Fig. 3. Weighted selection of samples. Samples from the similar patches are added to
the samples from the original texel. Additional samples are selected according to a
weighting criteria that is based on their maximum angular difference from the samples
of original texel.
this particular point itself into account [21]. In the following, we present a novel
spatio-temporal sampling scheme that reduces the risks of a bias in the BRDF
estimates by also taking into account dynamic reflectance samples from other
surface points with similar material properties.
4 Spatio-temporal Reflectance Sharing
By looking at its appearance from each camera view over time, we can generate
for each surface point, or equivalently, for each texel x a set of N appearance
samples
Dyx(x) = {Si | Si = (Ii, ˆ li, ˆvi), i ∈ {1, . . . , N}} (1)
Each sample Si stores a tuple of data comprising of the captured image intensity
Ii (from one of the cameras), the direction to the light source ˆ li, and the viewing
direction ˆvi. Please note that only if a point has been illuminated by exactly one
light source, a sample is generated. If a point is totally in shadow, illuminated
by two light sources, or not seen from the camera, no sample is created. Our
acquisition setup comprising of only 8 cameras and 2 light sources is comparably
simple and inexpensive. However, the fixed relative arrangement of cameras and
light sources may induce a bias in Dyx(x). There are two primary reasons for
this:
– Due to the fixed relative arrangement of cameras and light sources, each
surface point is only seen under a fixed number of half vector directions
ˆh = ˆ l + ˆv.
VII
Fig. 4. Texture-space layout of surface patches. Patches of same material are clustered
according to the average normal direction. For this illustration, patches of the same
material are colored in the same overall tone (e.g. blue for the shirt) but different
intensities.
– Even if the person performs a very expressive motion in the RES, samples lie
on “slices” of the hemispherical space of possible incoming light and outgoing
viewing directions.
Both of these factors possibly lead to BRDF estimates that may not generalize
well to lighting conditions that are very different to the acquisition setup.
By means of a novel spatio-temporal sampling strategy, called spatio-temporal
reflectance sharing, we can reduce the bias, Fig. 3. The guiding idea behind this
novel scheme is to use more than the samples Dyx(x) that have been measured
for the point x itself while the BRDF parameters for the point x are estimated.
The additional samples, combined in a set Dyxcompl (x), stem from other locations
on the surface that are made of similar material. These additional samples
have potentially been seen under different lighting and viewing directions than
the samples from Dyx(x) and can thus expand the sampling range. It is the
main challenge to incorporate these samples into the reflectance estimation at
x in a way that augments the generality of the measured BRDFs but does not
compromise the ability to capture spatial variation in surface appearance.
By explaining each step that is taken to draw samples for a particular surface
point x, we illustrate how we attack this challenge:
In a first step, the surface is clustered into patches of similar average normal
directions and same material, Fig. 4. Materials are clustered by means of
a simple k-means clustering using average diffuse colors [21]. The normal direction
ˆn of x defines the reference normal direction, Fig. 3a. Now, a list L of
patches consisting of the same material as x is generated. L is sorted according
to increasing angular deviation of average patch normal direction and reference
normal direction, Fig. 3b. Now, np many patches P0, ..., Pnp
are drawn from L
by choosing every lth list element. From each patch, a texel is selected at random,
resulting in a set of texels, T = xP0 , ..., xPnp . The set of texels T has been
VIII
selected in a way that maximizes the number of different surface orientations.
From the reflectance samples associated with texels in T , we now select a subset
Dyx compl (x) that maximizes the coverage of the 4D hemispherical space of
light and view directions. In order to decide which samples from T are potential
candidates for this set, we employ the following selection mechanism.
A weighting function δ(S1, S2) is applied that measures the difference of two
samples S1 = ( ˆ l1, ˆv1) and S2 = ( ˆ l2, ˆv2) in the 4D sample space as follows:
δ(S1, S2) = ∆( ˆ l1, ˆ l2) + ∆(ˆv1, ˆv2) (2)
where ∆ denotes the angular difference between two vectors. We employ δ to
select for each sample Sr in T its closest sample Sclosest in Dyx(x), i.e. the
sample for which ωSr = δ(Sr, Sclosest ) is minimal, Fig. 3d. Each sample Sr is
now weighted by ωSr . Only the ⌈αN⌉ samples from T with the highest weights
eventually find their way into Dyxcompl (x), Fig. 3e. The BRDF parameters
for x are estimated by taking all of the samples from Dyx(x) � Dyxcompl (x)
into account, Fig. 3f. Through experiments we have found out that a value
of α = 0.66 represents a good compromise between estimation robustness and
increase in computation time.
5 Results
We have tested our spatio-temporal reflectance sharing method on several input
sequences. The sequences of 150-300 frames cover two different human subjects
with different types of apparel. The method integrates seamlessly in the original
pipeline and no modifications of any kind are required for the rendering system.
We have verified both visually and quantitatively that our novel reflectance sampling
method leads to BRDF estimation that generalizes better to lighting conditions
different from the acquisition setup. Fig. 5 shows a side-by-side comparison
between the results obtained with and without spatio-temporal reflectance sharing.
Both human subjects are rendered under real world illumination using HDR
environment maps. Importance sampling is used to obtain direction light sources
that approximate the lighting from the static environment map [7]. Relightable
free-viewpoint videos can be rendered from arbitrary viewpoints and under arbitrary
lighting conditions at 6 fps if 16 approximating lights are employed. One
can see that with the exploitation of spatial coherence, more surface detail is
preserved under those lighting conditions which are strongly different from acquisition
setup. A small comparison video demonstrating only the relighting can
be seen here: http://www.mpi-inf.mpg.de/∼nahmed/Mirage2007.avi.
In addition to visual comparison, we also validated the method by comparing
the average peak-signal-to-noise-ratio with respect to input video stream obtained
under different lighting conditions. We have recorded one of our male test
subjects under two different calibrated lighting setups, henceforth termed LC A
and LC B. In each of the lighting setups, just one spot light has been employed.
The positions of the light sources in LC A and LC B are (angularly) approximately
45 ◦ apart with respect to the center of the scene. We reconstructed the
(a) (b)
(c) (d)
Fig. 5. Comparison of renditions under captured real-world illumination such as the
St Peter’s Basilica environment map (a),(b) and the Grace Cathedral environment
(c),(d) provided by Paul Debevec. One can see that compared to renditions obtained
without spatio-temporal reflectance sharing ((a) (c)), subtle surface details are much
better reproduced in the renditions obtained with spatio-temporal reflectance sharing
((b) (d)).
BRDF of the test subject under lighting setup LC B with and without our new
reflectance sampling. Subsequently, we calculated the PSNR with the ground
truth images of the person illuminated under setup LC A. Using our novel sampling
method, we have estimated surface reflectance using different percentages
of additional samples. For each case, we computed the PSNR with respect to the
ground truth. Fig. 6 shows the results that we obtained. Note that the graph of
IX
X
Fig. 6. PSNR values with respect to ground truth for different numbers of additional
samples Dyx compl (x).
the original method (green line) is constant over the increasing number of samples
just for the illustration purpose because it only considers the samples from
a single texel. With spatio-temporal reflectance sharing (red line) both results
are exactly the same in the beginning as no additional samples are considered,
but it can be seen that the PSNR improves as additional samples are taken into
account. We get a peak at around 30%-40% of additional samples. With the
inclusion of more samples the PSNR gradually decreases as the ever increasing
number of additional samples compromises the estimation of the reflectance’s
spatial variance. At maximum, we obtain a PSNR improvement of 0.75 dB. Although
we have performed the PSNR evaluation only for one sequence, we are
confident that for others it will exhibit similar results. This assumption is further
supported by the more compelling visual appearance obtained for all the other
test data that we have used.
Our approach is subject to a couple of limitations. We currently neglect interreflections
on the body. In the RES, they potentially play a role between the
wrinkles in clothing. For reflectance sharing, samples from the wrinkles can lead
to erroneous estimation. To prevent this effect from degrading the estimation
accuracy, we have taken care to minimize the number of wrinkles in the RES.
Sometimes, we observe small rendering artefacts due to undersampling (e.g. on
the underneath of the arms). However, we have verified that the application of
a RES sequence showing several rotation motions with different body postures
almost completely solves this problem. If even in this extended RES a pixel is
never seen by any of the cameras, we fill in reflectance properties from neighboring
regions in texture space.
Despite the limitations, our results show that spatio-temporal reflectance
sharing enable faithful estimation of dynamic reflectance models with only a
handful of cameras.
6 Conclusions
We have presented a spatio-temporal reflectance sharing method that reduces
the bias in BRDF estimation for dynamic scenes. Our algorithm exploits spatial
coherence by pooling samples of different surface location to robustify reflectance
estimation. In addition, it exploits temporal coherence by taking into
consideration samples from different steps of video. Despite the spatial-temporal
resampling, our algorithm is capable of reliably capturing spatially-varying reflectance
properties. By means of spatio-temporal reflectance sharing, we obtain
convincing 3D video renditions in real-time even under lighting conditions which
differ strongly from the acquisition setup.
7 Acknowledgements
This project has been supported by EU 3DTV NoE project No 511568.
References
1. Fausto Bernardini, Ioana M. Martin, and Holly Rushmeier. High-quality texture
reconstruction from multiple scans. IEEE TVCG, 7(4):318–332, 2001.
2. J. Carranza, C. Theobalt, M.A. Magnor, and H.-P. Seidel. Free-viewpoint video of
human actors. In Proc. of SIGGRAPH’03, pages 569–577, 2003.
3. Per Einarsson, Charles-Felix Chabert, Andrew Jones, Wan-Chun Ma, Bruce Lamond,
im Hawkins, Mark Bolas, Sebastian Sylwan, and Paul Debevec. Relighting
human locomotion with flowed reflectance fields. In Rendering Techniques, pages
183–194, 2006.
4. Athinodoros S. Georghiades. Recovering 3-d shape and reflectance from a small
number of photographs. In Eurographics Symposium on Rendering, pages 230–240,
2003.
5. D. Goldman, B. Curless, A. Hertzmann, and S. Seitz. Shape and spatially-varying
brdfs from photometric stereo. In Proc. of ICCV, pages 341–448, 2004.
6. Markus H. Gross, Stephan Würmlin, Martin Näf, Edouard Lamboray, Christian P.
Spagno, Andreas M. Kunz, Esther Koller-Meier, Tomás Svoboda, Luc J. Van Gool,
Silke Lang, Kai Strehlke, Andrew Vande Moere, and Oliver G. Staadt. blue-c: a
spatially immersive display and 3d video portal for telepresence. ACM Trans.
Graph. (Proc. of SIGGRAPH’03), 22(3):819–827, 2003.
7. Vlastimil Havran, Miloslaw Smyk, Grzegorz Krawczyk, Karol Myszkowski, and
Hans-Peter Seidel. Importance Sampling for Video Environment Maps. In ACM
SIGGRAPH 2005 Full Conference DVD-ROM, Los Angeles, USA, August 2005.
ACM SIGGRAPH, ACM. Sketches & Applications.
8. T. Hawkins, A. Wenger, C. Tchou, A. Gardner, F. Göransson, and P. Debevec.
Animatable facial reflectance fields. In Proc. of Eurographics Symposium on Rendering,
pages 309–319, 2004.
XI
XII
9. Takeo Kanade, Peter Rander, and P. J. Narayanan. Virtualized reality: Constructing
virtual worlds from real scenes. IEEE MultiMedia, 4(1):34–47, 1997.
10. E. Lafortune, S. Foo, K. Torrance, and D. Greenberg. ”non-linear approximation
of reflectance functions”. In Proceedings of SIGGRAPH’97, pages 117–126, August
1997.
11. Hendrik P. A. Lensch, Jan Kautz, Michael Goesele, Wolfgang Heidrich, and Hans-
Peter Seidel. Image-based reconstruction of spatial appearance and geometric
detail. ACM Transactions on Graphics, 22(2):27, 2003.
12. Ming Li, Hartmut Schirmacher, Marcus Magnor, and Hans-Peter Seidel. Combining
stereo and visual hull information for on-line reconstruction and rendering of
dynamic scenes. In Proc. of IEEE Multimedia and Signal Processing, pages 9–12,
2002.
13. T. Matsuyama and T. Takai. Generation, visualization, and editing of 3D video.
In Proc. of 1st International Symposium on 3D Data Processing Visualization and
Transmission (3DPVT’02), page 234ff, 2002.
14. W. Matusik, C. Buehler, R. Raskar, S.J. Gortler, and L. McMillan. Image-based
visual hulls. In Proceedings of ACM SIGGRAPH 00, pages 369–374, 2000.
15. W. Matusik, H. Pfister, M. Brand, and L. McMillan. A data-driven reflectance
model. ACM Trans. Graph. (Proc. SIGGRAPH’03), 22(3):759–769, 2003.
16. B.-T. Phong. Illumnation for computer generated pictures. Communications of
the ACM, pages 311–317, 1975.
17. H. Rushmeier, G. Taubin, and A. Guéziec. Applying Shape from Lighting Variation
to Bump Map Capture. In Eurographics Workshop on Rendering, pages 35–44,
June 1997.
18. Yoichi Sato, Mark D. Wheeler, and Katsushi Ikeuchi. Object Shape and Reflectance
Modeling from Observation. In Proc. of SIGGRAPH’97, pages 379–388, 1997.
19. J. Starck, G. Miller, and A. Hilton. Video-based character animation. In Proc.
ACM Symposium on Computer Animation, 2005.
20. C. Theobalt, J. Carranza, M. Magnor, and H.-P. Seidel. Combining 3d flow fields
with silhouette-based human motion capture for immersive video. Graphical Models,
66:333–351, September 2004.
21. Christian Theobalt, Naveed Ahmed, Hendrik P. A. Lensch, Marcus Magnor, and
Hans-Peter Seidel. Seeing people in different light - joint shape, motion and reflectance
capture. IEEE Transactions on Visualization and Computer Graphics,
to appear.
22. M. Waschbüsch, S. Würmlin, D. Cotting, F. Sadlo, and M. Gross. Scalable 3D
video of dynamic scenes. In Proc. of Pacific Graphics, pages 629–638, 2005.
23. A. Wenger, A. Gardner, C. Tchou, J. Unger, T. Hawkins, and P. Debevec. Performance
relighting and reflectance transformation with time-multiplexed illumination.
In ACM TOG (Proc. of SIGGRAPH’05), volume 24(3), pages 756–764,
2005.
24. S. Würmlin, E. Lamboray, O.G. Staadt, and M.H. Gross. 3d video recorder. In
Proc. of IEEE Pacific Graphics, pages 325–334, 2002.
25. Ruo Zhang, Ping-Sing Tsai, James Cryer, and Mubarak Shah. Shape from Shading:
A Survey. IEEE Trans. PAMI, 21(8):690–706, 1999.
26. Todd Zickler, Sebastian Enrique, Ravi Ramamoorthi, and Peter N. Belhumeur.
Reflectance sharing: Image-based rendering from a sparse set of images. In Proc.
of Eurographics Symposium on Rendering, pages 253–264, 2005.
27. C. Lawrence Zitnick, Sing Bing Kang, Matthew Uyttendaele, Simon Winder, and
Richard Szeliski. High-quality video view interpolation using a layered representation.
ACM TOC (Proc. SIGGRAPH’04), 23(3):600–608, 2004.
