This is the authors’ version of the work.

It is posted here by permission of Springer for

personal use. Not for redistribution. The final publication is available at link.springer.com.

Let There Be Color!

Large-Scale Texturing of 3D Reconstructions

Michael Waechter, Nils Moehrle, and Michael Goesele

TU Darmstadt

Abstract. 3D reconstruction pipelines using structure-from-motion and

multi-view stereo techniques are today able to reconstruct impressive,
large-scale geometry models from images but do not yield textured re-
sults. Current texture creation methods are unable to handle the com-
plexity and scale of these models. We therefore present the first compre-
hensive texturing framework for large-scale, real-world 3D reconstruc-
tions. Our method addresses most challenges occurring in such recon-
structions: the large number of input images, their drastically varying
properties such as image scale, (out-of-focus) blur, exposure variation,
and occluders (e.g., moving plants or pedestrians). Using the proposed
technique, we are able to texture datasets that are several orders of mag-
nitude larger and far more challenging than shown in related work.

1 Introduction
In the last decade, 3D reconstruction from images has made tremendous progress.
Camera calibration is now possible even on Internet photo collections [20] and for
city scale datasets [1]. There is a wealth of dense multi-view stereo reconstruction
algorithms, some also scaling to city level [7,8]. Realism is strongly increasing:
Most recently Shan et al . [18] presented large reconstructions which are hard to
distinguish from the input images if rendered at low resolution. Looking at the
output of state of the art reconstruction algorithms one notices, however, that
color information is still encoded as per-vertex color and therefore coupled to
mesh resolution. An important building block to make the reconstructed models
a convincing experience for end users while keeping their size manageable is
still missing: texture. Although textured models are common in the computer
graphics context, texturing 3D reconstructions from images is very challenging
due to illumination and exposure changes, non-rigid scene parts, unreconstructed
occluding objects and image scales that may vary by several orders of magnitudes
between close-up views and distant overview images.
So far, texture acquisition has not attracted nearly as much attention as
geometry acquisition: Current benchmarks such as the Middlebury multi-view
stereo benchmark [17] focus only on geometry and ignore appearance aspects.
Furukawa et al . [8] produce and render point clouds with very limited resolution,
which is especially apparent in close-ups. To texture the reconstructed geometry
2 M. Waechter, N. Moehrle, M. Goesele

Fig. 1. Left to right: Automatically textured model reconstructed from a set of images,
mesh close-up, and the same mesh rendered with texture.

Frahm et al . [7] use the mean of all images that observe it which yields insufficient
visual fidelity. Shan et al . [18] perform impressive work on estimating lighting
parameters per input image and per-vertex reflectance parameters. Still, they use
per-vertex colors and are therefore limited to the mesh resolution. Our texturing
abilities seem to be lagging behind those of geometry reconstruction.
While there exists a significant body of work on texturing (Section 2 gives a
detailed review) most authors focus on small, controlled datasets where the above
challenges do not need to be taken into account. Prominent exceptions handle
only specialized cases such as architectural scenes: Garcia-Dorado et al . [10] re-
construct and texture entire cities, but their method is specialized to the city
setting as it uses a 2.5D scene representation (building outlines plus estimated
elevation maps) and a sparse image set where each mesh face is visible in very
few views. Also, they are restricted to regular block city structures with planar
surfaces and treat buildings, ground, and building-ground transitions differently
during texturing. Sinha et al . [19] texture large 3D models with planar surfaces
(i.e., buildings) that have been created interactively using cues from structure-
from-motion on the input images. Since they only consider this planar case,
they can optimize each surface independently. In addition, they rely on user
interaction to mark occluding objects (e.g., trees) in order to ignore them dur-
ing texturing. Similarly, Tan et al . [22] propose an interactive texture mapping
approach for building façades. Stamos and Allen [21] operate on geometry data
acquired with time-of-flight scanners and therefore need to solve a different set
of problems including the integration of range and image data.
We argue that texture reconstruction is vitally important for creating realistic
models without increasing their geometric complexity. It should ideally be fully
automatic even for large-scale, real-world datasets. This is challenging due to
the properties of the input images as well as unavoidable imperfections in the
reconstructed geometry. Finally, a practical method should be efficient enough to
handle even large models in a reasonable time frame. In this paper we therefore
present the first unified texturing approach that handles large, realistic datasets
 Let There Be Color! — Large-Scale Texturing of 3D Reconstructions 3

reconstructed from images with a structure-from-motion plus multi-view stereo

pipeline. Our method fully automatically accounts for typical challenges inherent
in this setting and is efficient enough to texture real-world models with hundreds
of input images and tens of millions of triangles within less than two hours.

2 Related Work
Texturing a 3D model from multiple registered images is typically performed in
a two step approach: First, one needs to select which view(s) should be used to
texture each face yielding a preliminary texture. In the second step, this texture
is optimized for consistency to avoid seams between adjacent texture patches.

View Selection The literature can be divided into two main classes: Several
approaches select and blend multiple views per face to achieve a consistent tex-
ture across patch borders [5,13]. In contrast, many others texture each face with
exactly one view [9,10,15,23]. Sinha et al . [19] also select one view, but per texel
instead of per face. Some authors [2,6] propose hybrid approaches that generally
select a single view per face but blend close to texture patch borders.
Blending images causes problems in a multi-view stereo setting: First, if
camera parameters or the reconstructed geometry are slightly inaccurate, texture
patches may be misaligned at their borders, produce ghosting, and result in
strongly visible seams. This occurs also if the geometric model has a relatively
low resolution and does not perfectly represent the true object geometry. Second,
in realistic multi-view stereo datasets we often observe a strong difference in
image scale: The same face may cover less than one pixel in one view and several
thousand in another. If these views are blended, distant views blur out details
from close-ups. This can be alleviated by weighting the images to be blended [5]
or by blending in frequency space [2,6], but either way blending entails a quality
loss because the images are resampled into a common coordinate frame.
Callieri et al . [5] compute weights for blending as a product of masks indi-
cating the suitability of input image pixels for texturing with respect to angle,
proximity to the model, and proximity to depth discontinuities. They do however
not compute real textures but suggest the use of vertex colors in combination
with mesh subdivision. This contradicts the purpose of textures (high resolution
at low data cost) and is not feasible for large datasets or high-resolution images.
Similarly, Grammatikopoulos et al . [13] blend pixels based on angle and proxim-
ity to the model. A view-dependent texturing approach that also blends views
is Buehler et al .’s Lumigraph [4]. In contrast to the Lumigraph we construct a
global textured model and abstain from blending.
Lempitsky and Ivanov [15] select a single view per face based on a pairwise
Markov random field. Their data term judges the quality of views for textur-
ing while their smoothness term models the severity of seams between texture
patches. Based on this Allène et al . [2] and Gal et al . [9] proposed data terms
that incorporate additional effects compared to the basic data term. Since these
methods form the base for our technique, we describe them in Section 3.
4 M. Waechter, N. Moehrle, M. Goesele

Luminance
Patch 2

Patch 1 Patch 3

x x x
Before adjustment Velho, after adjust. Lempitsky, after adjust.

Fig. 2. Color adjustment (here in 1D). Left: Patch 2 is lighter than Patch 1 and 3, e.g.
due to different exposure. Center: Velho and Sossai [23] let the luminance transition
smoothly towards the seam’s mean. Right: Lempitsky and Ivanov [15] adjust globally.

Color Adjustment After view selection the resulting texture patches may
have strong color discontinuities due to exposure and illumination differences or
even different camera response curves. Thus, adjacent texture patches need to
be photometrically adjusted so that their seams become less noticeable.
This can be done either locally or globally. Velho and Sossai [23] (Figure 2
(center)) adjust locally by setting the color at a seam to the mean of the left and
right patch. They then use heat diffusion to achieve a smooth color transition
towards this mean, which noticeably lightens Patches 1 and 3 at their borders.
In contrast, Lempitsky and Ivanov [15] compute globally optimal luminance
correction terms that are added to the vertex luminances subject to two intuitive
constraints: After adjustment luminance differences at seams should be small and
the derivative of adjustments within a texture patch should be small. This allows
for a correction where Patch 2 is adjusted to the same level as Patch 1 and 3
(Figure 2 (right)) without visible meso- or large-scale luminance changes.

3 Assumptions and Base Method

Our method takes as input a set of (typically several hundred) images of a

scene that were registered using structure-from-motion [1,20]. Based on this the
scene geometry is reconstructed using any current multi-view stereo technique
(e.g., [7,8]) and further post-processed yielding a good (but not necessarily per-
fect) quality triangular mesh. This setting ensures that the images are registered
against the 3D reconstruction but also yields some inherent challenges: The
structure-from-motion camera parameters may not be perfectly accurate and
the reconstructed geometry may not represent the underlying scene perfectly.
Furthermore, the input images may exhibit strong illumination, exposure, and
scale differences and contain unreconstructed occluders such as pedestrians.
We now give an overview over how Lempitsky and Ivanov [15] and some
related algorithms work since our approach is based on their work. Section 4
describes the key changes made in our approach to handle the above challenges.
The initial step in the pipeline is to determine the visibility of faces in the
input images. Lempitsky and Ivanov then compute a labeling l that assigns a
view li to be used as texture for each mesh face Fi using a pairwise Markov
 Let There Be Color! — Large-Scale Texturing of 3D Reconstructions 5

random field energy formulation (we use a simpler notation here):

X X
E(l) = Edata (Fi , li ) + Esmooth (Fi , Fj , li , lj ) (1)
Fi ∈Faces (Fi ,Fj )∈Edges

The data term Edata prefers “good” views for texturing a face. The smoothness
term Esmooth minimizes seam (i.e., edges between faces textured with different
images) visibility. E(l) is minimized with graph cuts and alpha expansion [3].
As data term the base method uses the angle between viewing direction and
face normal. This is, however, insufficient for our datasets as it chooses images
regardless of their proximity to the object, their resolution or their out-of-focus
blur. Allène et al . [2] project a face into a view and use the projection’s size as
data term. This accounts for view proximity, angle and image resolution. Similar
to this are the Lumigraph’s [4] view blending weights, which account for the very
same effects. However, neither Allène nor the Lumigraph account for out-of-focus
blur: In a close-up the faces closest to the camera have a large projection area
and are preferred by Allène’s data term or the Lumigraph weights but they may
not be in focus and lead to a blurry texture. Thus, Gal et al . [9] use the gradient
magnitude of the image integrated over the face’s projection. This term is large
if the projection area is large (close, orthogonal images with a high resolution)
or the gradient magnitude is large (in-focus images).
Gal et al . also introduce two additional degrees of freedom into the data term:
They allow images to be translated by up to 64 pixels in x- or y-direction to
minimize seam visibility. While this may improve the alignment of neighboring
patches, we abstain from this because it only considers seam visibility and does
not explain the input data. In a rendering of such a model a texture patch would
have an offset compared to its source image. Also, these additional degrees of
freedom may increase the computational complexity such that the optimization
becomes infeasible for realistic dataset sizes.
Lempitsky and Ivanov’s smoothness term is the difference between the tex-
ture to a seam’s left and right side integrated over the seam. This should prefer
seams in regions where cameras are accurately registered or where misalignments
are unnoticeable because the texture is smooth. We found, that computation of
the seam error integrals is a computational bottleneck and cannot be precom-
puted due to the prohibitively large number of combinations. Furthermore, it
favors distant or low-resolution views since a blurry texture produces smaller
seam errors, an issue that does not occur in their datasets.
After obtaining a labeling from minimizing Equation 1, the patch colors are
adjusted as follows: First, it must be ensured that each mesh vertex belongs to
exactly one texture patch. Therefore each vertex on a seam is duplicated into
two vertices: Vertex vleft belonging to the patch to the left and vright belonging
to the patch to the right of the seam.1 Now each vertex v has a unique color fv
1
In the following, we only consider the case where seam vertices belong to n = 2
patches. For n > 2 we create n copies of the vertex and optimize all pairs of those
copies jointly, yielding a correction factor per vertex and patch.
6 M. Waechter, N. Moehrle, M. Goesele

before adjustment. Then, an additive correction gv is computed for each vertex,

by minimizing the following expression (we use a simpler notation for clarity):
X
2 1 X
2
argmin fvleft + gvleft − (fvright + gvright ) + gvi − gvj (2)
g λ
v (split into vi ,vj are ad-
vleft and vright ) jacent and in
lies on a seam the same patch

The first term ensures that the adjusted color to a seam’s left (fvleft + gvleft ) and
its right (fvright + gvright ) are as similar as possible. The second term minimizes
adjustment differences between adjacent vertices within the same texture patch.
This favors adjustments that are as gradual as possible within a texture patch.
After finding optimal gv for all vertices the corrections for each texel are inter-
polated from the gv of its surrounding vertices using barycentric coordinates.
Finally, the corrections are added to the input images, the texture patches are
packed into texture atlases, and texture coordinates are attached to the vertices.

4 Large-Scale Texturing Approach

Following the base method we now explain our approach, focusing on the key
novel aspects introduced to handle the challenges of realistic 3D reconstructions.

4.1 Preprocessing

We determine face visibility for all combinations of views and faces by first per-
forming back face and view frustum culling, before actually checking for occlu-
sions. For the latter we employ a standard library [11] to compute intersections
between the input model and viewing rays from the camera center to the trian-
gle under question. This is more accurate than using rendering as, e.g., done by
Callieri et al . [5], and has no relevant negative impact on performance. We then
precompute the data terms for Equation 1 for all remaining face-view combina-
tions since they are used multiple times during optimization, remain constant,
and fit into memory (the table has O(#faces·#views) entries and is very sparse).

4.2 View Selection

Our view selection follows the structure of the base algorithm, i.e., we obtain a
labeling such as the one in Figure 3 (left) by optimizing Equation 1 with graph
cuts and alpha expansion [3]. We, however, replace the base algorithm’s data and
smoothness terms and augment the data term with a photo-consistency check.

Data Term For Rthe reasons described in Section 3 we choose Gal et al .’s [9] data
term Edata = − φ(Fi ,li ) k∇(Ili (p))k2 dp. We compute the gradient magnitude
k∇(Ili )k2 of the image into which face Fi is projected with a Sobel operator
and sum over all pixels of the gradient magnitude image within Fi ’s projection
 Let There Be Color! — Large-Scale Texturing of 3D Reconstructions 7

φ(Fi , li ). If the projection contains less than one pixel we sample the gradient
magnitude at the projection’s centroid and multiply it with the projection area.
The data term’s preference for large gradient magnitudes entails an important
problem that Gal et al . do not account for because it does not occur in their
controlled datasets: If a view contains an occluder such as a pedestrian that has
not been reconstructed and can thus not be detected by the visibility check,
this view should not be chosen for texturing the faces behind that occluder.
Unfortunately this happens frequently with the gradient magnitude term (e.g.
in Figure 9) because occluders such as pedestrians or leaves often feature a larger
gradient magnitude than their background, e.g., a relatively uniform wall. We
therefore introduce an additional step to ensure photo-consistency of the texture.

Photo-Consistency Check We assume that for a specific face the majority of

views see the correct color. A minority may see wrong colors (i.e., an occluder)
and those are much less correlated. Based on this assumption Sinha et al . [19] and
Grammatikopoulos et al . [13] use mean or median colors to reject inconsistent
views. This is not sufficient, as we show in Section 5. Instead we use a slightly
modified mean-shift algorithm consisting of the following steps:

1. Compute the face projection’s mean color ci for each view i in which the face
is visible.
2. Declare all views seeing the face as inliers.
3. Compute mean µ and covariance matrix Σ of all inliers’ mean color ci .

4. Evaluate a multi-variate Gaussian function exp − 21 (ci − µ)T Σ −1 (ci − µ) for
each view in which the face is visible.
5. Clear the inlier list and insert all views whose function value is above a thresh-
old (we use 6 · 10−3 ).
6. Repeat 3.–5. for 10 iterations or until all entries of Σ drop below 10−5 , the
inversion of Σ becomes unstable, or the number of inliers drops below 4.

We obtain a list of photo-consistent views for each face and multiply a penalty
on all other views’ data terms to prevent their selection.
Note, that using the median instead of the mean does not work on very
small query sets because for 3D vectors the marginal median is usually not a
member of the query set so that too many views are purged. Not shifting the
mean does not work in practice because the initial mean is often quite far away
from the inliers’ mean (see Section 5 for an example). Sinha et al . [19] therefore
additionally allow the user to interactively mark regions that should not be used
for texturing, a step which we explicitly want to avoid.

Smoothness Term As discussed above, Lempitsky and Ivanov’s smoothness

term is a major performance bottleneck and counteracts our data term’s pref-
erence for close-up views. We propose a smoothness term based on the Potts
model: Esmooth = [li 6= lj ] ([·] is the Iverson bracket). This also prefers compact
patches without favoring distant views and is extremely fast to compute.
8 M. Waechter, N. Moehrle, M. Goesele

Fig. 4. Left: A mesh. Vertex v1 is adjacent to both texture patches (red and blue).
Its color is looked up as a weighted average over samples on the edges v0 v1 and v1 v2 .
Right: Sample weights transition from 1 to 0 as distance to v1 grows.

4.3 Color Adjustment

Models obtained from the view selection

phase (e.g., Figure 3 (right)) contain many
color discontinuities between patches. These
need to be adjusted to minimize seam visibil-
ity. We use an improved version of the base
method’s global adjustment, followed by a
local adjustment with Poisson editing [16].

Global Adjustment A serious problem

with Lempitsky and Ivanov’s color adjust-
Fig. 3. Left: A mesh’s labeling. Each
ment is that fvleft and fvright in Equation 2 color represents a different label, i.e.
are only evaluated at a single location: the input image. Right: The textured re-
vertex v’s projection into the two images ad- sult with visible luminance differ-
jacent to the seam. If there are even small ences between patches.
registration errors (which there always are),
both projections do not correspond to exactly the same spot on the real object.
Also, if both images have a different scale the looked up pixels span a different
footprint in 3D. This may be irrelevant in controlled lab datasets, but in real-
istic multi-view stereo datasets the lookups from effectively different points or
footprints mislead the global adjustment and produce artifacts.

Color Lookup Support Region We alleviate this problem by not only looking
up a vertex’ color value at the vertex projection but along all adjacent seam
edges, as illustrated by Figure 4: Vertex v1 is on the seam between the red and
the blue patch. We evaluate its color in the red patch, fv1,red , by averaging color
samples from the red image along the two edges v0 v1 and v1 v2 . On each edge
we draw twice as many samples as the edge length in pixels. When averaging
the samples we weight them according to Figure 4 (right): The sample weight
is 1 on v1 and decreases linearly with a sample’s distance to v1 . (The reasoning
behind this is that after optimization of Equation 2 the computed correction
gv1,red is applied to the texels using barycentric coordinates. Along the seam the
barycentric coordinates form the transition from 1 to 0.) We obtain average colors
 Let There Be Color! — Large-Scale Texturing of 3D Reconstructions 9

for the edges v0 v1 and v1 v2 , which we average weighted with the edge lengths
to obtain fv1,red . Similarly we obtain fv1,blue and insert both into Equation 2.
For optimization, Equation 2 can now be written in matrix form as

2 2
kAg − f k2 + kΓgk2 = gT (AT A + ΓT Γ)g − 2f T Ag + f T f (3)

f is a vector with the stacked fvleft −fvright from Equation 2. A and Γ are sparse
matrices containing ±1 entries to pick the correct gvleft , gvright , gvi and gvj from
g. Equation 3 is a quadratic form in g and AT A+ΓT Γ is very sparse, symmetric,
2 2
and positive semidefinite (because ∀z: zT (AT A + ΓT Γ)z = kAzk2 + kΓzk2 ≥ 0).
We minimize it with respect to g with Eigen’s [14] conjugate gradient (CG)
implementation and stop CG when krk2 / AT f 2 < 10−5 (r is the residual),
which typically requires < 200 iterations even for large datasets. Due to auto-
matic white balancing, different camera response curves and different light colors
(noon vs. sunset vs. artificial light) between images it is not sufficient to only
optimize the luminance channel. We thus optimize all three channels in parallel.

Poisson Editing Even with the above support regions Lempitsky and Iva-
nov’s global adjustment does not eliminate all visible seams, see an example
in Figure 11 (bottom row, center). Thus, subsequent to global adjustment we
additionally perform local Poisson image editing [16]. Gal et al . [9] do this as well,
but in a way that makes the computation prohibitively expensive: They Poisson
edit complete texture patches, which results in huge linear systems (with > 107
variables for the largest patches in our datasets).
We thus restrict the Poisson editing of a patch to
a 20 pixel wide border strip (shown in light blue in
Figure 5). We use this strip’s outer rim (Fig. 5, dark
blue) and inner rim (Fig. 5, red) as Poisson equation
boundary conditions: We fix each outer rim pixel’s
Fig. 5. A texture patch
value to the mean of the pixel’s color in the image has a border strip (light
assigned to the patch and the image assigned to the blue) with an outer (dark
neighboring patch. Each inner rim pixel’s value is fixed blue) and inner rim (red).
to its current color. If the patch is too small, we omit
the inner rim. The Poisson equation’s guidance field is the strip’s Laplacian.
For all patches we solve the resulting linear systems in parallel with Ei-
gen’s [14] SparseLU factorization. For each patch we only compute the factor-
ization once and reuse it for all color channels because the system’s matrix stays
the same. Adjusting only strips is considerably more time and memory efficient
than adjusting whole patches. Note, that this local adjustment is a much weaker
form of the case shown in Figure 2 (center) because patch colors have been ad-
justed globally beforehand. Also note, that we do not mix two images’ Laplacians
and therefore still avoid blending.
10 M. Waechter, N. Moehrle, M. Goesele

dataset # views # mesh faces image resolution

Statue 334 4.9 million 5616×3744
City Wall 561 8.2 million 2000×1500
Castle Ruin 287 20.3 million 5616×3744

Statue calculating data costs

view selection
City Wall global seam leveling
Castle Ruin local seam leveling
building OBJ model
10 20 30 40 50 60 70 80
runtime in minutes (wall clock)

Fig. 6. Summary of the three datasets used in the evaluation and runtime of the
individual parts of the algorithm.

5 Evaluation
We now evaluate the proposed approach using three datasets (Statue, City Wall,
Castle Ruin) of varying complexity, see Figure 6 for a summary of their prop-
erties. Even our largest dataset (Castle Ruin, Figure 7) can be textured within
less than 80 min on a modern machine with two 8 core Xeon E5-2650v2 CPUs
and 128 GB of memory. Main computational bottlenecks are the data cost com-
putation whose complexity is linear in the number of views and the number
of pixels per view, and the graph cut-based view selection which is linear in
the number of mesh faces and the number of views. When altering the above
datasets properties, e.g., by simplifying the mesh, we found that the theoretical
complexities fit closely in practice. The data term computation is already fully
parallelized but the graph cut cannot be parallelized easily due to the graph’s ir-
regular structure. Still, this is in stark contrast to other methods which may yield
theoretically optimal results while requiring a tremendous amount of computa-
tion. E.g., Goldluecke et al . compute for several hours (partially on the GPU)
to texture small models with 36 views within a super-resolution framework [12].
In the following sections, we evaluate the individual components of our ap-
proach and compare them to related work.

Data Term and Photo-Consistency Check As argued previously, the data

term must take the content of the candidate images and not only their geomet-
ric configuration into account. Allène et al .’s data term [2] computes a face’s
projection area and thus selects the image in the center of Figure 8 for texturing
the statue’s arm, even though this image is not focused on the arm but on its
background. Gal et al .’s data term [9] favors instead large gradient magnitudes
and large projection areas. It thus uses a different image for the arm, yielding a
much crisper result as shown in Figure 8 (right).
A failure case of Gal’s data term are datasets with little surface texture detail.
For example the Middlebury Dino [17] is based on a uniform white plaster statue,
all views have the same distance to the scene center and out-of-focus blur is not
 Let There Be Color! — Large-Scale Texturing of 3D Reconstructions 11

Fig. 7. Some texturing results of the Castle Ruin dataset.

Fig. 8. Left: Statue’s arm textured with Allène’s [2] data term. Center: Detail from
the image used for texturing the arm in the left image, exhibiting out-of-focus blur.
Right: The arm textured with Gal’s [9] data term. (Best viewed on screen.)

an issue. Thus, Gal’s data term accounts for effects that do not occur in this
dataset. We found that the data term instead overfits to artifacts such as shadows
and the color adjustment is incapable of fixing the resulting artifacts.
Another drawback of Gal’s data term is, that it frequently uses occluders
to texture the background (as shown in Figure 9 (top right)) if the occluder
contains more high-frequency image content than the correct texture.
In Figure 9 (bottom left) we show the photo-consistency check at work for one
exemplary face. This face has many outliers (red), i.e. views seeing an occluder
instead of the correct face content. The gray ellipse marks all views that our
check classifies as inliers. Only one view is misclassified. Note the black path
taken by the shifted mean: The starting point on the path’s bottom left is the
algorithm’s initial mean. Sinha et al . [19] and Grammatikopoulos et al . [13]
use it without iterating in mean-shift fashion. It is, however, far off from the
true inlier mean and if we used a fixed window around it many views would be
misclassified. If we used all views’ covariance matrix instead of a fixed window
all views except for the bottommost data point would be classified as inliers.
Our approach only misclassifies the one outlier closest to the inlier set and is
able to remove the pedestrian (see Figure 9 (bottom right)).
12 M. Waechter, N. Moehrle, M. Goesele

Fig. 9. Top left: Input mesh. Top right: Pedestrian used as texture for the ground. Bot-
tom left: Photo-consistency check visualization for a single face. Red and blue crosses
are the face’s mean color (here we show only the Y and Cb component) in different
images. Bottom right: Mesh textured with photo-consistency check.

Our proposed photo-consistency check may fail, e.g. if a face is seen by very
few views. In this case there is little image evidence for what is an inlier or an
outlier and the classification becomes inaccurate. However, in our experience this
is often fixed by the view selection’s smoothness term: If view X is erroneously
classified as outlier for face A, but not for face B, it may still be optimal to choose
it for texturing both A and B if the data penalty for face A is smaller than the
avoided smoothness penalty between A and B. This can also fail, especially in
border cases where there are not enough neighbor faces that impose a smoothness
penalty, such as in the bottom right corner of the bottom right image in Figure 9.

Smoothness Term Lempitsky and Ivanov use the error on a seam integrated
along the seam as smoothness term for the view selection. The result is a model
where blurry texture patches selected from distant views occur frequently since
they make seams less noticeable (see Figure 10 (left)). Using the Potts model
instead yields a considerably sharper result (see Figure 10 (right)).

Color Adjustment The global color adjustment used in the base method [15]
operates on a per-vertex base. It fails if a seam pixel does not project onto the
 Let There Be Color! — Large-Scale Texturing of 3D Reconstructions 13

Fig. 10. Left: Statue’s shoulder with Lempitsky and Ivanov’s smoothness term.
Right: Resulting model with the Potts model smoothness term. (Best viewed on screen.)

Fig. 11. Top row: Unadjusted mesh with wireframe, detail of input image used for
upper patch, and detail of input image used for lower patch. Bottom row: Artifact from
global color adjustment caused by texture misalignment, result of global adjustment
with support region, and result after Poisson editing.

Fig. 12. Top row: One of the input images. Bottom row: Reconstructed City Wall color
adjusted without support region and without Poisson editing, and the same reconstruc-
tion adjusted with support region and Poisson editing. (Best viewed on screen.)
14 M. Waechter, N. Moehrle, M. Goesele

same 3D position in the images on both sides of its seam. This can occur for
various reasons: Camera misalignment, 3D reconstruction inaccuracies or a dif-
ferent 3D footprint size of the looked up pixels if the projected resolution differs
in both images. In Figure 11 (top left) the vertex on the right side of the letter
’J’ projects into the letter itself to the seam’s top and into the lighter back-
ground to the seam’s bottom. The result after global adjustment (bottom left)
exhibits a strong color artifact. We alleviate this using support regions during
sampling (bottom center). The remaining visible seams are fixed with Poisson
editing as shown in Figure 11 (bottom right). Poisson editing does obviously not
correct texture alignment problems but disguises most seams well so that they
are virtually invisible unless the model is inspected from a very close distance.
These color adjustment problems in Lempitsky and Ivanov’s method occur
frequently when mesh resolution is much lower than image resolution (which
is the key application of texturing). Vertex projections’ colors are in this case
much less representative for the vertices’ surroundings. Figure 12 shows such
an example: The result without support region and Poisson editing (bottom
left) exhibits several strong color artifacts (white arrows) when compared to a
photograph of the area (top). The result with support region and Poisson editing
(bottom right) significantly improves quality.

6 Conclusions and Future Work

Applying textures to reconstructed models is one of the keys to realism. Surpris-
ingly, this topic has been neglected in recent years and the state of the art is still
to reconstruct models with per-vertex colors. We therefore presented the first
comprehensive texturing framework for large 3D reconstructions with registered
images. Based on existing work, we make several key changes that have a large
impact: Large-scale, real-world geometry reconstructions can now be enriched
with high-quality textures which will significantly increase the realism of recon-
structed models. Typical effects occurring frequently in these datasets, namely
inaccurate camera parameters and geometry, lighting and exposure variation,
image scale variation, out-of-focus blur and unreconstructed occluders are han-
dled automatically and efficiently. In fact, we can texture meshes with more than
20 million faces using close to 300 images in less than 80 minutes.
Avenues for future work include improving the efficiency of our approach and
accelerating the computational bottlenecks. In addition, we plan to parallelize
on a coarser scale to texture huge scenes on a compute cluster in the spirit of
Agarwal et al .’s [1] structure-from-motion system.

Acknowledgements We are grateful for financial support by the Intel Visual

Computing Institute through the project RealityScan. Also, we thank Victor
Lempitsky and Denis Ivanov for providing the code for their paper.

Source Code This project’s source code can be downloaded from the project
web page www.gris.informatik.tu-darmstadt.de/projects/mvs-texturing.
 Let There Be Color! — Large-Scale Texturing of 3D Reconstructions 15

References
1. Agarwal, S., Snavely, N., Simon, I., Seitz, S.M., Szeliski, R.: Building Rome in a
day. In: ICCV (2009) 1, 4, 14
2. Allène, C., Pons, J.P., Keriven, R.: Seamless image-based texture atlases using
multi-band blending. In: ICPR (2008) 3, 5, 10, 11
3. Boykov, Y., Veksler, O., Zabih, R.: Fast approximate energy minimization via
graph cuts. PAMI 23 (2001) 5, 6
4. Buehler, C., Bosse, M., McMillan, L., Gortler, S., Cohen, M.: Unstructured lumi-
graph rendering. In: SIGGRAPH (2001) 3, 5
5. Callieri, M., Cignoni, P., Corsini, M., Scopigno, R.: Masked photo blending: Map-
ping dense photographic dataset on high-resolution sampled 3D models. Computers
& Graphics 32 (2008) 3, 6
6. Chen, Z., Zhou, J., Chen, Y., Wang, G.: 3D texture mapping in multi-view recon-
struction. In: Advances in Visual Computing, LNCS, vol. 7431 (2012) 3
7. Frahm, J.M., Fite-Georgel, P., Gallup, D., Johnson, T., Raguram, R., Wu, C.,
Jen, Y.H., Dunn, E., Clipp, B., Lazebnik, S., Pollefeys, M.: Building Rome on a
cloudless day. In: ECCV (2010) 1, 2, 4
8. Furukawa, Y., Curless, B., Seitz, S.M., Szeliski, R.: Towards Internet-scale multi-
view stereo. In: CVPR (2010) 1, 4
9. Gal, R., Wexler, Y., Ofek, E., Hoppe, H., Cohen-Or, D.: Seamless montage for
texturing models. Computer Graphics Forum 29 (2010) 3, 5, 6, 9, 10, 11
10. Garcia-Dorado, I., Demir, I., Aliaga, D.G.: Automatic urban modeling using vol-
umetric reconstruction with surface graph cuts. Computers & Graphics 37 (2013)
2, 3
11. Geva, A.: ColDet 3D collision detection. sourceforge.net/projects/coldet 6
12. Goldluecke, B., Aubry, M., Kolev, K., Cremers, D.: A super-resolution framework
for high-accuracy multiview reconstruction. IJCV (2013) 10
13. Grammatikopoulos, L., Kalisperakis, I., Karras, G., Petsa, E.: Automatic multi-
view texture mapping of 3D surface projections. In: 3D-ARCH (2007) 3, 7, 11
14. Guennebaud, G., Jacob, B., others: Eigen v3. eigen.tuxfamily.org (2010) 9
15. Lempitsky, V., Ivanov, D.: Seamless mosaicing of image-based texture maps. In:
CVPR (2007) 3, 4, 12
16. Pérez, P., Gangnet, M., Blake, A.: Poisson image editing. ACM Transactions on
Graphics 22(3) (2003) 8, 9
17. Seitz, S.M., Curless, B., Diebel, J., Scharstein, D., Szeliski, R.: A comparison and
evaluation of multi-view stereo reconstruction algorithms. In: CVPR (2006) 1, 10
18. Shan, Q., Adams, R., Curless, B., Furukawa, Y., Seitz, S.M.: The visual Turing
test for scene reconstruction. In: 3DV (2013) 1, 2
19. Sinha, S.N., Steedly, D., Szeliski, R., Agrawala, M., Pollefeys, M.: Interactive 3D
architectural modeling from unordered photo collections. In: SIGGRAPH Asia
(2008) 2, 3, 7, 11
20. Snavely, N., Seitz, S.M., Szeliski, R.: Photo tourism: Exploring photo collections
in 3D. In: SIGGRAPH (2006) 1, 4
21. Stamos, I., Allen, P.K.: Geometry and texture recovery of scenes of large scale.
Computer Vision and Image Understanding 88(2) (2002) 2
22. Tan, Y., Kwoh, L., Ong, S.: Large scale texture mapping of building facades.
The International Archives of the Photogrammetry, Remote Sensing and Spatial
Information Sciences 37 (2008) 2
23. Velho, L., Sossai Jr., J.: Projective texture atlas construction for 3D photography.
The Visual Computer 23 (2007) 3, 4