Iterative Color Equalization for Increased

Applicability of Structured Light Reconstruction

Torben Fetzer

, Gerd Reis

and Didier Stricker

1,2

Department of Computer Science, University of Kaiserslautern, Germany

Department Augmented Vision, DFKI GmbH, Germany

Keywords:

Structured Light, 3D-Reconstruction, Color Equalization, Photometric Compensation.

Abstract:

The ﬁeld of 3D reconstruction is one of the most important areas in computer vision. It is not only of the-

oretical importance, but is also increasingly used in practice, be it in reverse engineering, quality control or

robotics. A distinction is made between active and passive methods, depending on whether they are based

on active interactions with the object or not. Due to the accuracy and density of the reconstructions obtained,

the structured light approach, whenever applicable, is often the method of choice for industrial applications.

Nevertheless, it is an active approach which, depending on material properties or coloration, can lead to prob-

lems and fail in certain situations. In this paper, a method based on the standard structured light approach

is presented that signiﬁcantly reduces the inﬂuence of the color of a scanned object. It improves the results

obtained by repeated application in terms of accuracy, robustness and general applicability. Especially in

high-precision reconstruction of small structures or high-contrast colored and specular objects, the technique

shows its greatest potential. The advanced method requires neither pre-calibrated cameras or projectors nor

information about the equipment. It is easy to implement and can be applied to any existing scanning setup.

1 INTRODUCTION

Modern structured light systems typically consist of

a projector that illuminates the scene through several

fringe patterns and at least one camera that captures

the lit scene. There are several techniques based on

binary patterns, pseudo-random patterns, and ﬁnally

phase-shifted structured light known as state of the

art. The calculation of phase images for each camera

from the patterns projected ensures continuous sur-

face encoding. This allows dense correspondence to

be determined between all views, including the pro-

jector. Based on the resulting high-quality correspon-

dences, modern auto-calibration techniques ((Zhang,

1998), (Fetzer et al., 2019), (Lourakis and Deriche,

2000), (Hartley and Zisserman, 2003)) ensure that

camera matrices can be estimated robustly, ready for

triangulation of point clouds. In this way, a calibra-

tion of the devices is generated that achieves highest

accuracy on the object’s surface. This in turn leads to

high-precision reconstructions and is a superiority of

this method over other scanning techniques.

This work was funded by the project MARMORBILD

(03VP00293) of the German Federal Ministry of Education

and Research (BMBF).

In addition to the strengths of structured light,

there is a number of disadvantages that should by

no means be neglected. The basis of the process is

the visibility of the projections on the object’s sur-

face. In this context, transparent, mirroring and spec-

ular scenes should be mentioned above all. Even ob-

jects whose texture contains both highly absorbent

and highly reﬂective areas can cause problems. In

many cases such scenes lead to inaccuracies and often

to a complete failure of the method. With this work

we want to eliminate or at least reduce some of these

problems in a way that can be easily adopted to exist-

ing scan setups. We signiﬁcantly increase the appli-

cation area of structured light reconstruction without

additional hardware requirements.

According to the state of the art, shifted sinusoidal

fringe patterns are used which sum up to zero. There-

fore, the encoding method should be invariant to the

object’s texturing. The following reﬂection properties

are unfortunately neglected in this idealized model:

• Different materials cause different light reﬂection.

• Different colors reﬂect the light in different ways.

Therefore, in many cases the appearance of the pro-

jected fringe patterns can change into the impercep-

Fetzer, T., Reis, G. and Stricker, D.

Iterative Color Equalization for Increased Applicability of Structured Light Reconstruction.

DOI: 10.5220/0008879607250732

In Proceedings of the 15th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2020) - Volume 4: VISAPP, pages

725-732

ISBN: 978-989-758-402-2; ISSN: 2184-4321

725

Figure 1: Recorded section of a 10-euro note (top left) and

corresponding point cloud, reconstructed by structured light

(top right). The enlarged area (below) shows errors in the

estimated depth, solely caused by the object’s coloration.

tible. Most reconstructions of objects made of stan-

dard material are not very strongly affected and time-

consuming procedures, to always treat this behavior,

would certainly be overdone. Nevertheless, this ef-

fect is clearly noticeable in high-precision reconstruc-

tions and it is worth taking a closer look. To illustrate

this, Figure 1 shows a captured area of a 10-euro note

and a point cloud thereof reconstructed with struc-

tured light. The enlargement of the point cloud clearly

shows the erroneous depth caused exclusively by the

texturing of the object. The combination of the pro-

jection patterns with an inverse texture signiﬁcantly

reduces this effect.

2 RELATED WORK

Modern phase-shifted structured light systems with

digital devices were introduced in (Zhang and Huang,

2004). In order to shorten the acquisition time,

(Zhang and Huang, 2006) and before (Zhang et al.,

2002) introduced methods that use color-coded pat-

terns. (Sansoni and Redaelli, 2005) and later (Yang

et al., 2017) introduced single shot structured light

techniques, where the multiple shifted fringe im-

ages were coded by carrier waves and combined to

one pattern. (Donlic et al., 2015) and (Petkovi

et al., 2016) presented single-shot structured light ap-

proaches, based on de bruijn color sequences. All

methods to shorten the acquisition time reduced the

quality of reconstructions signiﬁcantly. In (Zhang and

Yau, 2008) a setup with two cameras is presented

to signiﬁcantly increase the scan quality. When us-

ing several cameras (ideally the same camera model)

there are many advantages with regard to calibration,

gamma correction of the recorded scenes and the re-

sulting quality of the reconstructions. Although, the

approach based on this technique has become state

of the art in many areas, there are several scenarios,

where this approach is not applicable.

Extensive research has been carried out to im-

prove the applicability to general situations. In order

to cope with strong ambient lighting such as sunlight,

(Gupta et al., 2013) introduced a possibility of com-

pensation by sequentially allocating a given energy

budget to several sections. (Nayar et al., 2006) and

later (O’Toole et al., 2014), (O’Toole et al., 2012) pre-

sented ways to split direct and indirect light paths, en-

abling the reconstruction of mirroring, reﬂecting and

light emitting objects and even scanning through dust.

Unfortunately, this requires expensive hardware, the

process is error-prone and requires high-precision cal-

ibration of camera pixels to a DLP panel, making

it difﬁcult to use for practical applications. (Wein-

mann et al., 2011) increased the range of application

of structured light but using additional devices.

3 INVERSE TEXTURE

First of all, we want to investigate the inﬂuence of

a projection to the captured image of a scene. We

then present a procedure that modulates the projec-

tion in such a way that as many points as possible in

the scene have an equivalent inﬂuence to the captured

image. This procedure equalizes the appearance of an

object iteratively and converges after only a few steps.

The projector-camera correspondences are generated

using the structured light approach and are reliably

cleaned of erroneous phases with the help of a newly

introduced simple masking method.

Let I be a captured image of a scene that was il-

luminated by a projection T . Schreiber and Bruning

(Malacara, 2007) described the physical inﬂuence of

T on image I. Accordingly, the captured image can

be approximated as the composition

I = I

+ I

◦ T, (1)

of the average intensity I

and the scene intensity I

moduled by the projected texture T . Thereby, ◦ de-

notes the element-wise multiplication operator.

In order to minimize the inﬂuence of the coloring

of an object to its image, projection texture T has to be

estimated so that it balances the object colors as much

as possible. Therefore, the following minimization

problem has to be solved:

argmin

∑

i j

−

= argmin

∑

i j

+ I

i j

−

(2)

While I

i j

denotes the pixels of image I and

I the op-

timal common color value of the equalized pixels.

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

726

(a) Camera-Projector Correspondences (b) Masking Erroneous Phase (c) Inverse Texture Projection

Figure 2: Examples of sinusoidal fringe patterns and thus illuminated scenes captured by a camera (a, top). Horizontal and

vertical phases of the projector and the camera calculated from the fringe images (a, bottom). Texture image calculated

from averaging captures (4) and error image calculated from (9) (b, left). Projection mask clustered from error image and

correspondingly masked phase images (b, right). Textures T projected onto an exemplary scene: Gray projection and inverse

texture calculated with Algorithm 1 (c, bottom). Correspondingly captured scene without and with color correction (c, top).

Projecting T on the scene approximates an uncolored

grayish scene as visualized in Figure 2 (c).

Theoretically, an optimal texture T solution (2)

can be calculated explicitly from at least two captured

images. However, this is not recommended from a

practical point of view. This is due to the following

phenomena, which cannot be considered in the un-

derlying model (1):

• Color cross-talk between the different color chan-

nels of projector and camera.

• Scattered light from one pixel to others are not

considered, since it only takes into account direct

pixel to pixel correspondences.

• Specular or absorbing materials, leading to

clipped values in the captured image, due to a lim-

ited dynamic range.

3.1 Iterative Color Equalization

To solve problem (2), we propose an iterative method

that is robust against the irregularities listed above.

We combine alternate updates of the projected texture

T and the equalized target value

I with a logarithmic

search that uses the limited range of projector pixel

values (8-bit in[0,255]). In this way, a stable conver-

gence of the process is achieved after a few iterations.

Logarithmic Search. We assume that we have

direct correspondences between camera and

projector pixels that are given by a mapping

P : (x

, y

) → (x

, y

) that assigns a corresponding

projector pixel (x

, y

) to each image pixel (x

, y

Therewith, any projection pixel T

i j

in the captured

scene is known to be located at P (T

i j

) in the projector

input image. Since these are usually limited to

8-bit color values in three color channels, we have a

discrete search range for texture values of T that can

be effectively used by a binary search. Moreover,

projecting light is a monotonous procedure, therefore

increasing values of T

i j

lead to increasing values of

i j

. In order to implement and exploit this knowledge,

the values of the projection P (T ) are adjusted via a

logarithmic search until the error (2) of the resulting

image I = I

+ I

◦ T to the equalization value

I for

all pixels is minimal.

Equalization Value. Minimizing energy (2) with a

ﬁxed texture T leads to an optimal equalization value

I =

∑

i j

(3)

given by the mean of image I

M×N

Alternating updates of the inverse texture T and

the equalization value

I with adjusted increments lead

to Algorithm 1, which already converges after 7 it-

erations in case of standard 8-bit projective devices.

Since the texture update is pixel-wise independent,

the individual iterations can be implemented efﬁ-

ciently. Figure 2 (c) shows a scene balanced in this

way. Each iteration was applied separately to the dif-

ferent color channels to intercept the color cross-talk.

3.2 Camera-projector Correspondences

In order to apply Algorithm 1, a reliable mapping

P (·) as described in Section 3.1 is required. A rec-

ommended approach for determining close point cor-

respondences between projector image and camera

image is the structured light approach introduced in

(Zhang and Huang, 2004). The projection of phase-

shifted sine waves (Figure 2, (a), top) allows the cal-

culation of phase images encoding the scene through

the projection (Figure 2, (a), bottom). This usually re-

quires phase unwrapping methods like (Bioucas-Dias

Iterative Color Equalization for Increased Applicability of Structured Light Reconstruction

727

and Valadao, 2007) or (An et al., 2016) to encode gen-

erated wrapped phases. The phase information can be

used to determine point correspondences of projector

and camera pixels. Errors in the underlying phase in-

formation can be caused by

• Overexposed or underexposed areas in the scene

where the projected fringes are not visible.

• Regions in the scene that are visible to the camera

but not to the projector.

• Shadows cast by the illuminated object.

Masking Erroneous Phase Values. For the success

of the proposed color equalization method it is a pre-

requisite to have an accurate mapping P available. No

false correspondences should be used that would sig-

niﬁcantly falsify the result. For this purpose, incorrect

phase information should be masked out beforehand.

Defective phase regions, which are calculated using

standard phase shift approaches (Malacara, 2007),

are usually much more noisy than correctly coded

ones. Therefore, gradient based ﬁlters are typically

used to mask out erroneous regions. However, these

approaches are not sufﬁciently accurate for the pre-

sented application. Due to the gradient dependency,

edges are falsely masked out, which runs counter to

the later goal of reconstructing highly accurate small

structures. In order to create an appropriate mask-

ing, a simple method is presented that provides much

more accurate results. Given is the basic property of

phase-shifted sinusoidal patterns:

∑

n=1

∑

m=1

= I

+ 0.5I

. (4)

The sum of the phase-shifted sine waves results in

zero, which neglects their inﬂuence. This means that

the applied illumination of the scene in sum is equiv-

alent to a uniform grey projection. Let I

and I

denote the captured scenes, illuminated by respective

sine patterns P

and P

, deﬁned by

(i, j ) = sin



2π j

2π(n − 1)



(5)

(i, j ) = sin



2πi

2π(m − 1)



(6)

i =1, ..., h, j = 1, ..., w, n = 1, ..., N, m = 1, ..., M

with F

and F

being the number of horizontal and

vertical fringes over the projection and w and h the

projector’s image width and height in pixels.

If the scene was captured without illumination, I

is already given and so I

can be estimated by

∑

n=1

∑

m=1

− 2I

. (7)

Finally, we can calculate an error E of the horizontal

and vertical phase values Φ

and Φ

to the captured

images I

and I

E =

∑

n=1



sin



2π(n−1)



−

−I



(8)

∑

m=1



sin



2π(m−1)



−

−I



Figure 2 (b, top left) shows an example of a texture

computed from image means (4) and (b, bottom left)

the respective error image E from (9). This error re-

liably indicates the quality of the phase values in re-

lation to all captured images of the scene. Since erro-

neous phase values produce much higher errors than

correct ones, a high-quality mask can be generated by

applying k-Means Clustering to error image E. Note

that bi-clustering can be efﬁciently implemented in

O(MNlog(MN)). Figure 2 (b, top right) shows the

ﬁnal mask as a result of k-means clustering applied

with two clusters. Figure 2 (b, bottom right) shows

the ﬁnal masked phases.

Algorithm 1: Iterative Color Equalization.

Input: Camera-projector correspondences P .

1: Initialize projection texture P (T

(0)

i j

) = 128 ∀i j.

2: Project pattern T

(0)

i j

and capture lit scene I

(0)

for z = 1, ..., 7

(z−1)

∑

i j

(z−1)

i j

4: P (T

(z)

i j

) =











P (T

(z−1)

i j

) + 2

7−z

, I

(z−1)

i j

(z−1)

P (T

(z−1)

i j

) − 2

7−z

, I

(z−1)

i j

(z−1)

P (T

i j

), else

5: Project pattern T

(z)

and capture resulting scene I

(z)

end

Output: Inverse texture T .

4 INVERSELY TEXTURING

STRUCTURED LIGHT (ITSL)

In order to neglect color inﬂuences on the geometry

estimation by the structured light approach, we com-

bine the inverse texture calculated by Algorithm 1

with the fringe images P

and P

from (5) and (6).

Instead of the normal patterns we project the Inversely

Texturing Structured Light Patterns (ITSLP)

= P

◦ P (

T ), n = 1, ..., N (9)

= P

◦ P (

T ), m = 1, ..., M. (10)

We lift values of the inverse texture close to zero to

avoid that no fringes are projected in these regions

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

728

Figure 3: Captured scene of a zebra to visualize the improvements of several iterations of ITSL to color equalization. Normal

image (left) and equalized captures after one, two and three iterations (second left to right).

after multiplication:

i j

(

T · T

i j

, if T

i j

> 0.05 · max(T )

0.05 · max(T ) , else

(11)

In the process, masked areas are also coded after sev-

eral iterations of the approach.

An important feature of ITSLP is that they fulﬁll

the basic property for fringes of a structured light sys-

tem, as mentioned in (4). In the new case we have for

every scene

∑

n=1

= I

∑

n=1

−1

) ◦

T ◦ I

(12)

= I

+ 0.5(

T ◦ I

) ≈ I

+ 0.5(T ◦ I

). (13)

This is equivalent to usual structured light patterns

being projected onto a grey scene without color in-

ﬂuence. Figure 4 (top left) shows ITSLP, computed

by (11) and the patterns projected onto the scene

(top right). For further visualization Figure 4 (bot-

tom) shows the scenes after averaging (4) the standard

structured light patterns (left) and ITSLP (right).

Each iteration of ITSLP increases the quality of

the reconstructions. Regions with incorrect phase in-

formation of the ﬁrst iteration can be corrected by

multiple iterations of ITSL.

Figure 4: Examples of inversely texturing fringe images

projected by the projector (top left) and illuminated scene

(top right). Textures calculated by (7) from the fringe im-

ages (bottom).

5 EVALUATION

In order to evaluate the usefulness of the method pre-

sented, we carry out some quantitative and qualitative

tests. First the performance of the color equalization

from Algorithm 1 is examined. In particular, the be-

havior after several iterations of ITSL is investigated.

Subsequently, we show the advantages and the impor-

tant practical beneﬁts of ITSLP to structured light re-

construction. In several scenarios, in which the stan-

dard structured light approach usually fails, the ben-

eﬁts, which arise from the new, improved approach,

become clear.

5.1 Inverse Projection Texture

Figure 6 (left) shows a captured image of a standard-

ized X-Rite ColorChecker normally used for color

calibration. It consists of 24 calibrated colors that

well cover the entire visual color spectrum. Figure 6

(right) shows the captured checkerboard after an itera-

tion of Algorithm 1. The method applied to this object

demonstrates its behavior in case of very large color

differences in an object’s texture. It is clearly visible

that some color patches (dark red, yellow) cannot be

equalized completely. The reason for this is that light

can only be projected but not removed. If the red,

green or blue component is already in an area above

the mean value of the equalization, it cannot be re-

duced by any projection. To be exact, the gamut of

the camera does not lie in the gamut generated by the

projector. Nevertheless, equalization results of this

quality lead to a signiﬁcant improvement in the reﬂec-

tive properties of an object. Moreover, since grayscale

images are sufﬁcient for reconstruction, monochrome

cameras can also be used. Errors that may occur dur-

ing color equalization due to inconsistent gamuts are

negligible in this case. Multiple iterations of ITSL,

further improve the quality of equalization. Figure 6

shows the behavior of the Root-Mean-Squared-Error

(RMSE), referred to the mean value (3), for an iter-

ative application. While after the ﬁrst iteration the

RMSE decreases by more than 90%, the following it-

Iterative Color Equalization for Increased Applicability of Structured Light Reconstruction

729

Figure 5: Flat colored checkerboard: Initial reconstructed point cloud (top left), normal texture (top middle) and equalized

texture after one iteration of ITSL (top right). Enlarged marked area of point cloud (bottom left) and point clouds before and

after two iterations of the proposed approach (bottom second left to right).

erations reduce this error only slightly in this case.

However, these minor changes can lead to a dra-

matic improvement of reconstructions. In particular,

the reﬂection behaviour at edges of strongly contrast-

ing areas can be signiﬁcantly improved after a few

iterations, due to lower radiation. For visual evalua-

tion, Figure 3 shows several iterations of the method

applied to a ﬁgurine of a zebra. The coloration of

this object contains maximally strong edges in the

transitions from black to white areas. Due to better

approximated phase values, every iteration improves

the equalization quality. Note that limited projector

brightness and stray light between pixels are the rea-

sons why it is impossible to achieve complete equal-

ization in an extreme scenario like the one shown in

Figure 3. But within the scope of the possibilities, the

result obtained here is by far sufﬁcient to yield signiﬁ-

cant improvements in reconstruction, as demonstrated

later on.

Figure 6: Captured image of a X-Rite ColorChecker before

(top left) and after one iteration of color equalization with

Algorithm 1 (top right). RMSE of color equalization for

several iterations of ITSLP (bottom).

5.2 Inversely Texturing Structured

Light

In order to demonstrate the advantages of the pre-

sented method in the context of 3D reconstruction,

we will apply it in the following exemplary scenarios.

We show quantitative and qualitative improvements in

the important case of high-precision reconstruction of

very small structures. We then apply it to objects with

high-contrast staining, such as the one in Figure 3,

and qualitatively investigate the behavior after several

iterations of the method. Finally, the chances arising

from the method in context of specular and reﬂecting

objects are shown by an exemplary reconstruction of

a shiny metal sphere.

5.2.1 High-precision Reconstruction

Different colors of an object’s texture reﬂect different

wavelengths of light. These speciﬁc properties, de-

pending on the object coloration, cause projected pat-

terns in structured light applications to be reﬂected in

slightly different ways. Therefore, depending on the

coloration of an object, slightly different depth values

are estimated. Usually, this effect is very small, com-

pared to the geometry of an object, and can therefore

be neglected. However, in high-precision 3D recon-

structions, as for example encountered in quality con-

trol setups, this problem has considerable effects and

signiﬁcantly distorts the results, especially in high-

contrast regions.

In order to evaluate the usefulness of the proce-

dure in relation to this problem, we apply it to a

ﬂat checkerboard with patches of different colors, as

shown in Figure 5. The checkerboard is absolutely ﬂat

and any differences in the depth of the reconstruction

are errors due to the different colors of the patches.

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

730

Figure 7: Average Euclidean depth error in µm of recon-

structed point cloud for several iterations of ITSL.

Figure 5 (top left) shows the reconstructed point cloud

resulting from the standard structured light approach.

For further visualization, Figure 5 (top center, right)

shows the captured scene before and after color equal-

ization. To demonstrate the problem more clearly,

Figure 5 (bottom left) shows an enlarged version of

the marked area.

Figure 5 (bottom, second left) shows the point

cloud of the enlarged region without texture informa-

tion, but colored by the Euclidean error with respect

to the ﬂat ground truth. Finally, Figure 5 (bottom,

second right and right) shows the reconstructed re-

gions after one and two iterations of ITSL. To make a

qualitative evaluation possible, the depth value of the

checkerboards is enhanced by a factor of 3 for visual-

ization. This clearly shows the improvements of the

method presented.

Besides the qualitative evaluation, we demonstrate

the beneﬁt by a quantitative error measurement. Fig-

ure 7 shows the behavior of the average depth error

of the reconstructed point with respect to the ground

truth. Multiple iterations improve the quantitative er-

ror continuously. Nevertheless, since the improve-

ments are in the range of µm, one should decide

whether the improvement of accuracy of the recon-

struction justiﬁes higher additional expenses in the

speciﬁc case.

5.2.2 High-contrast Colored Objects

Another important ﬁeld of application of the method

is the reconstruction of objects with extremely un-

favourable colouring. The statue of a zebra from Fig-

ure 3 is treated as an example. Due to the very bright

and very dark areas there is no camera setting that

allows a complete encoding of the surface with struc-

tured light. Figure 8 (left) shows the result of standard

structured light with a rather short exposure time. The

white areas of the zebra are well reconstructed, while

the black areas are underexposed and not encoded by

the patterns. Conversely, Figure 8 (middle) shows the

reconstructions at a higher exposure time, which al-

lows the reconstruction of the black areas, but over-

exposes the white regions. Finally, Figure 8 (right)

shows the result of ITSL. Already one iteration can

solve the problem caused by the color contrasts and

allows the reconstruction of the entire surface. Fur-

Figure 8: Reconstructed point clouds of a zebra statue. Top

row and bottom row show the same point clouds with and

without color information. Left and middle columns show

results of standard structured light in case of over and under-

exposed images. Right shows ITSL handling the proposed

problems.

ther iterations improve the quality slightly, but they

should again be weighed according to the beneﬁt and

the recording time spent.

5.2.3 Specular Objects

As last and most important ﬁeld of application of the

method, we show its beneﬁts in structured light recon-

struction of specular objects. We apply the method

to a specular metal sphere that strongly reﬂects the

light emitted by the projector. The resulting high-

lighted areas are overexposed and cannot be encoded

by the projected patterns. This effect cannot be com-

pletely avoided, but it can be greatly reduced. Figure

9 shows the reconstructed point clouds with standard

structured light (left) and after one and two iterations

of ITSL (center and right).

The highlighted points do not only depend on the

projector’s position relative to the object, but also on

the camera perspective. To illustrate that the overex-

posed point is different for different camera positions,

the point clouds in Figure 9 are triangulated from a

pair of cameras instead of a camera-projector pair.

Therefore, there are two independent faulty holes in

the reconstructed point clouds from Figure 9. Finally,

we would like to point out that this property means,

that the defective area of the phase of one camera

is most likely correctly encoded in the other camera.

This can be used to improve reconstructions and to

make them invariant to reﬂective objects in multiple

camera structured light setups, as they are typical for

practical applications.

6 CONCLUSION

In this paper, we have presented a method that greatly

expands the practical scope of structured light re-

construction. Typical scenarios, in which the stan-

dard approach usually fails, are now treatable. Un-

Iterative Color Equalization for Increased Applicability of Structured Light Reconstruction

731

Figure 9: Triangulated point clouds of a specular sphere,

captured from two views. Standard structured light (left)

and results of ITSL after one and two iterations (middle and

right). The top row shows the textured point clouds, while

the bottom row visualizes the sole geometry.

favorably colored and reﬂective objects present no or

signiﬁcantly fewer problems. In the area of high-

precision reconstruction, a signiﬁcant leap in accu-

racy is achieved. However, the iterative character of

the method also increases the recording time. Sev-

eral iterations increase the accuracy, but should be

weighed against the additional time required. There-

fore, the approach is designed so that it can be eas-

ily built on existing setup and applied or omitted as

needed.

REFERENCES

An, Y., Hyun, J.-S., and Zhang, S. (2016). Pixel-

wise absolute phase unwrapping using geometric con-

straints of structured light system. Optics express,

24(16):18445–18459.

Bioucas-Dias, J. M. and Valadao, G. (2007). Phase unwrap-

ping via graph cuts. IEEE Transactions on Image pro-

cessing, 16(3):698–709.

Donlic, M., Petkovic, T., and Pribanic, T. (2015). 3d surface

proﬁlometry using phase shifting of de bruijn pattern.

In Proceedings of the IEEE International Conference

on Computer Vision, pages 963–971.

Fetzer, T., Reis, G., and Stricker, D. (2019). Robust auto-

calibration for practical scanning setups from epipolar

and trifocal relations. In 2019 16th International Con-

ference on Machine Vision Applications (MVA), pages

1–6. IEEE.

Gupta, M., Yin, Q., and Nayar, S. K. (2013). Structured

light in sunlight. In Proceedings of the IEEE Inter-

national Conference on Computer Vision, pages 545–

552.

Hartley, R. and Zisserman, A. (2003). Multiple view geom-

etry in computer vision. Cambridge university press.

Lourakis, M. I. and Deriche, R. (2000). Camera self-

calibration using the kruppa equations and the svd of

the fundamental matrix: The case of varying intrinsic

parameters.

Malacara, D. (2007). Optical shop testing. John Wiley &

Sons.

Nayar, S. K., Krishnan, G., Grossberg, M. D., and Raskar,

R. (2006). Fast separation of direct and global com-

ponents of a scene using high frequency illumination.

In ACM Transactions on Graphics (TOG), volume 25,

pages 935–944. ACM.

O’Toole, M., Mather, J., and Kutulakos, K. N. (2014).

3d shape and indirect appearance by structured light

transport. In Proceedings of the IEEE Conference

on Computer Vision and Pattern Recognition, pages

3246–3253.

O’Toole, M., Raskar, R., and Kutulakos, K. N. (2012).

Primal-dual coding to probe light transport. ACM

Trans. Graph., 31(4):39–1.

Petkovi

c, T., Pribani

c, T., and Ðonli

c, M. (2016). Single-

shot dense 3d reconstruction using self-equalizing de

bruijn sequence. IEEE Transactions on Image Pro-

cessing, 25(11):5131–5144.

Sansoni, G. and Redaelli, E. (2005). A 3d vision system

based on one-shot projection and phase demodulation

for fast proﬁlometry. Measurement Science and Tech-

nology, 16(5):1109.

Weinmann, M., Schwartz, C., Ruiters, R., and Klein,

R. (2011). A multi-camera, multi-projector super-

resolution framework for structured light. In 2011

International Conference on 3D Imaging, Modeling,

Processing, Visualization and Transmission, pages

397–404. IEEE.

Yang, L., Li, F., Xiong, Z., Shi, G., Niu, Y., and

Li, R. (2017). Single-shot dense depth sensing

with frequency-division multiplexing fringe projec-

tion. Journal of Visual Communication and Image

Representation, 46:139–149.

Zhang, L., Curless, B., and Seitz, S. M. (2002). Rapid shape

acquisition using color structured light and multi-pass

dynamic programming. In Proceedings. First Inter-

national Symposium on 3D Data Processing Visual-

ization and Transmission, pages 24–36. IEEE.

Zhang, S. and Huang, P. (2004). High-resolution, real-time

3d shape acquisition. In 2004 Conference on Com-

puter Vision and Pattern Recognition Workshop, pages

28–28. IEEE.

Zhang, S. and Huang, P. S. (2006). High-resolution, real-

time three-dimensional shape measurement. Optical

Engineering, 45(12):123601.

Zhang, S. and Yau, S.-T. (2008). Three-dimensional shape

measurement using a structured light system with dual

cameras. Optical Engineering, 47(1):013604.

Zhang, Z. (1998). Determining the epipolar geometry and

its uncertainty: A review. International journal of

computer vision, 27(2):161–195.

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

732