You are on page 1of 8

Exp Fluids (2008) 45:549–556

DOI 10.1007/s00348-008-0521-5

RESEARCH ARTICLE

Volume self-calibration for 3D particle image velocimetry


B. Wieneke

Received: 13 December 2007 / Revised: 25 March 2008 / Accepted: 16 May 2008 / Published online: 13 June 2008
Ó Springer-Verlag 2008

Abstract Planar self-calibration methods have become 1 Introduction


standard for stereo PIV to correct misalignments between
laser light sheet and calibration plane. Computing cross- Standard in situ calibration methods for stereo PIV and
correlation between images from camera 1 and 2 taken at volumetric velocimetry techniques employ an accurately
the same time, non-zero disparity vectors indicate rota- manufactured planar calibration plate with known marks
tional and translational misalignments relative to the translated in a controlled way through the illuminated sheet
coordinate system defined by a calibration plate. This or volume. With a dual-plane 3D calibration plate a single
approach works well for thin light sheets but fails for recording is sufficient for a thin sheet eliminating the need
extended volumes recorded in 3D-PTV or tomographic for a translation stage. Alternatively, one can use multiple
PIV experiments. Here it is primarily necessary to correct recordings of a planar calibration plate at arbitrary posi-
calibration errors leading to triangulation errors in 3D-PTV tions and process by camera pinhole model and bundle
or in degraded tomographic volume reconstruction. adjustment fit, which is a common procedure in computer
Tomographic PIV requires calibration accuracies of a vision. The outcome of the calibration is a mapping func-
fraction of a pixel throughout the complete volume, which tion M(X, Y, Z) from space to camera pixel coordinates
is difficult to achieve experimentally. A new volumetric (x, y):
self-calibration technique has been developed based on the
ðxi ; yi Þ ¼ Mi ðX; Y; Z Þ ð1Þ
computation of the 3D position of matching particles by
triangulation as in 3D-PTV. The residual triangulation for all cameras i. Details concerning calibration procedures
error (‘disparity’) is then used to correct the mapping and functional forms of the mapping function can be found
functions for all cameras. A statistical clustering method in the literature (Soloff et al. 1997; Prasad 2000; Calluaud
suitable for dense particle images has been implemented to and David 2004; Scarano et al. 2005). For stereo PIV there
find correct disparity map peaks from true particle matches. are basically two types of calibration errors.
Disparity maps from multiple recordings are summed for First the mapping function M itself can be erroneous due
better statistics. This self-calibration scheme has been to inaccurate calibration plates, inaccurate plate movement,
validated using several tomographic PIV experiments mechanical instabilities, optical distortions not accounted
improving the vector quality significantly. The relevance for by the mapping function and other error sources. Even
for other 3D velocimetry methods is discussed. with great care the maximum errors are typically above
0.5–2 pixel in some region of the mapped volume. This is
usually not a problem for stereo PIV, because the 2C-dis-
placement vectors are still computed accurately for both
cameras, only for the 3C-reconstruction the two vectors are
taken from slightly different world positions. Apart from
B. Wieneke (&) the case of very high velocity gradients a calibration error
LaVision GmbH, Anna-Vandenhoeck-Ring 19,
37081 Gottingen, Germany of 1 pixel is not relevant compared to other common error
e-mail: bwieneke@lavision.de sources.

123
550 Exp Fluids (2008) 45:549–556

More significant for stereo PIV is the second type of first summarized followed by a detailed explanation of
error resulting from a translational and rotational mis- each step.
alignment between the coordinate system defined by the The first step is localizing 2D particle positions in the
calibration plate position (at z = 0) compared to the actual image of each camera and finding corresponding 3D par-
light sheet position. Typical disparity errors are in the ticle positions by matching and triangulation which is
range of 5–20 pixel. Advanced ‘self-calibration’ or ‘mis- essentially the first part of the 3D-PTV algorithm as shown
alignment correction’ methods have been developed in Fig. 1 for three cameras. One should note that 3D-PTV
correcting such a misalignment using the actual recorded and tomographic PIV is usually done with four cameras.
particle images (Willert 1997; Coudert and Schon 2001; The 2D-particle positions (x, y) of camera i of a single
Scarano et al. 2005; Wieneke 2005). This is done by taking particle in space are triangulated into a best-fit world
images from camera 1 and 2 recorded at the same time, position (X, Y, Z). The optimization criterion is usually

then dewarping them to the world plane z = 0, which done by minimizing the sum of the distances x0i ; y0i 
corresponds to the position of the calibration plate, fol- ðxi ; yi Þj where ðx0i ; y0i Þ is the projection of (X, Y, Z) back to
lowed by cross-correlation. The resulting 2D disparity image of camera i (Hartley and Sturm 1994). For a perfect
vector field should be zero, that is, the two images are mapping function the line of sights of all cameras intersect
exactly matched. Non-zero disparity vectors can be used in in a single world point resulting in a zero disparity di of
an appropriate way to shift and rotate the coordinate system  
di ¼ dix ; diy ¼ x0i ; y0i  ðxi ; yi Þ ð2Þ
such that the new z = 0 plane coincides with the actual light
sheet position. Collecting disparities di for all particles throughout the
Even large displacements can be detected and corrected volume is then used to correct the mapping functions M to
eliminating the need to accurately align the calibration M0
plate with the light sheet in the first place. One should note Mi0 ðX; Y; Z Þ ¼ Mi ðX; Y; Z Þ  di ðX; Y; Z Þ ð3Þ
that the disparity correlation peak is actually an extended
streak with a length proportional to light sheet thickness. with an appropriate validation and interpolation of the
The correction method works well for thin Gaussian light sparse disparities di for each particle into a continuous field
sheets, but fails for thicker sheets and volumes since the or a similar method described below in more detail.
correlation streak becomes too long, too weak and frag- This basic principle works well with sparsely seeded
mented, even showing multiple peaks from non-Gaussian 3D-PTV images with, for example, 1,000–5,000 particles
intensity profiles. in each image, where the matching of corresponding par-
For scanning light sheet stereo PIV (Brücker 1996; Hori ticle images is done with a high confidence level ([95%)
and Sakakibara 2004) this self-calibration method can be (Maas 1996). But it does not work without further refine-
used to detect the exact position for all light sheets with a ment for densely seeded images from tomographic PIV.
single rough volume calibration before. Due to inaccura- Here the particle seeding density is much higher with
cies of the mechanical scanning mechanism such a typically [50,000 particles in each image equivalent to
calculation might be required for each recorded image. particle densities per pixel in the images of up to 0.1 ppp
For 3D-PTV and especially for tomographic PIV with (particles per pixel) similar to standard planar PIV. Ideally
extended volumes the first type of error relating to inac-
curate mapping functions is more relevant. The
tomographic reconstruction step requires that each voxel
position in space is mapped to a camera pixel position with
an error less than 0.4 pixel (Elsinga et al. 2006), preferably
less than 0.1 pixel. Therefore a volumetric self-calibration
technique is needed, which is capable of correcting the
mapping function throughout the volume. This is described
in the next section, followed by an experimental verifica-
tion using tomographic PIV showing significant
improvement of the vector field quality.

2 Method

The basic principle of correcting mapping function errors Fig. 1 Residual triangulation disparities di ðX; Y; Z Þ ¼ ðxi ; yi Þ 
in a volume using the actual recorded particle images is ðx0i ; y0i Þ due to calibration errors

123
Exp Fluids (2008) 45:549–556 551

one would record first a number of images with less correction. In addition given such high densities the
seeding density, but this is often inconvenient or even not probability of overlapping particles is high which leads to
possible. erroneous 2D and 3D particle positions and noisy
Figure 2 shows the procedure to match particle images disparities.
corresponding to the same particle in space. For each The solution is to apply an appropriate clustering tech-
particle in the image of camera 1 one finds particles located nique and using not all particles but only the brightest, for
within a stripe defined by the uncertainty emax along the example, 10%. Immediately the number of ghost particles
length Lz which is the projection of the line of sight of is reduced by nearly several orders of magnitude with the
camera 1 through the illuminated volume from z1 to z2 added advantage that even in case of overlapping particles
viewed by camera 2 (‘epipolar line’). For the combination the bright particles are much less influenced by another
of particle images for camera 1 and 2 a 3D particle position overlapping weak particle. Usually image preprocessing
(X, Y, Z) is computed by triangulation. Then for each including global and local intensity renormalization is
particle in the stripe it is checked if in the image of camera applied to the four camera images to subtract background
3 and 4 there is a corresponding particle within ±emax intensities and to correct particle intensity differences due
around the projection of (X, Y, Z) onto camera 3 and 4 to (e.g., forward–backward) laser scattering differences. A
image. The allowed uncertainty emax must be larger than single threshold for all camera images can then be used so
the largest expected calibration error. For 3D-PTV with that the group of particles above a certain threshold is about
low seeding density and four cameras one rarely detects the same in all camera images.
any incorrect ghost particles. The ratio R of ghost particles The complete self-calibration procedure can be divided
to true particles is a strong function of the seeding density into the following steps:
nppp and allowed error emax:
a. Measuring 2D particle positions in all camera images.
R ¼ 8  Lz  e5max  n3ppp ð4Þ b. Determining possible 3D particle positions by
Typical conditions with, for example, Lz = 200 pixel, triangulation.
nppp = 0.05 ppp and allowed error emax = 4 pixel lead to c. Subdividing the complete illuminated volume into nx 
R = 205 times more ghost particles than true ones. Obvi- ny  nz sub-volumes.
ously it is then very difficult to identify the true particles d. Plotting the disparities of all particles in a sub-volume
and their disparity to be used for the mapping function as a 2D disparity map.

Fig. 2 Particle triangulation


procedure

123
552 Exp Fluids (2008) 45:549–556

e. Summing disparity maps for each sub-volume and which is then used in Eq. 2. Of course, for the reference
all cameras over many recordings to get better camera the disparity is zero. Whether one should use a
statistics. reference camera deserves further investigation. Tomo-
f. For each disparity map detecting highest disparity peak graphic experiments so far have been mostly indifferent,
which is the most probable disparity vector di(X,Y,Z). with a few working better with one or the other method.
g. Validating and optionally smoothing of the 3D2C Having collected possible 3D particles in space one now
disparity vector field di. needs to separate disparities from true accurate 2D/3D
h. Correcting mapping functions Mi according to Eq. 3 particle positions from ghost particles and inaccurate ones
for all cameras i. due to overlapping particles. There are many suitable
i. Repeating the complete procedure as a check and for clustering techniques available. The one used here is based
further improvement. on displaying the disparities for each sub-volume in a 2D
disparity map as a histogram with each measurement point
Step a As described above a suitable threshold is plotted as a small 2D Gauss blob. For example, the dis-
selected to only use the brightest, for example, 10% of parity map might be of size of, for example, ±30 pixel
all 2D particle images. Particle fitting itself is done in a corresponding to a maximum allowed disparity emax of ±3
standard way with a 5 9 5 2D Gauss fit, reduced to a pixel, and each disparity di is plotted at position (dix, diy) as
3 9 3 Gauss fit if a neighboring particles gets too close, a 2D Gauss blob with a constant width of 2 pixel (0.2 pixel
that is, if an outer rim pixel is brighter than the inner in units of disparity error).
3 9 3 core, falling back to a dual x and y 1D Gauss fit The width of the Gauss curves provides sub-pixel
which always work. For isolated particles the position accuracy for the position of the highest disparity peak (step
accuracy is typically below 0.1 pixel depending on f) similar to correlation peaks in standard PIV. It should be
optical conditions and particle size etc. made larger in case of few detected particles. The height of
Step b The triangulation procedure computes all possible the Gauss curve is set to 1 for all particles. Alternatively,
particles in space with a triangulation uncertainty di one could use the particle intensity of the 2D peak finder to
(Eq. 2) smaller than emax as shown in Fig. 2. account for the light intensity variations across the volume.
Step c The complete illuminated volume is subdivided Such a display of the disparity allows a good visual
into nx  ny  nz sub-volumes (e.g., 5 9 5 9 5) in order to inspection of the data quality. This is shown in Fig. 3 for a
collect sufficient 3D particles in each sub-volumes. real tomographic PIV experiment. One can guess that the
Step d For each sub-volume and all 3D particles within true disparity peak is at the top-right, but there are not yet
the sub-volume a small 2D Gauss blob is plotted at the enough particles for good statistics.
disparity position (dix, diy) in a square of, for example,
50 9 50 pixels corresponding to the allowed maximum Step e Disparity maps from multiple recordings are
error ±emax. For each camera i a separate disparity map summed for better statistics. Most often a single image
is generated. does not contain sufficient particles to add-up to a sharp

The computation of the disparity di can be done in


slightly different ways. The first step is always to use the
particle positions (xi, yi) in the images of camera i to
compute a best-fit world position (X, Y, Z) by triangulation
(Hartley and Sturm 1994) as in step b together with the
residual triangulation error di as the difference between
(xi0 , yi0 ) = Mi (X, Y, Z) and (xi, yi) (Eq. 2). One has an extra
degree of freedom which allows one to use one camera as a
reference assuming that the mapping function for this
camera is perfect thus correcting only the other mapping
functions. This is done by taking Z from the original tri-
angulation and computing an Xref and Yref by inverting the
mapping function for the reference camera:
1
ðXref ; Yref Þ ¼ Mref ðxref ; yref ; Z Þ ð5Þ
and using Xref and Yref for computing (xi0 , yi0 ) for the other
cameras:

x0i ; y0i ¼ Mi ðXref ; Yref ; Z Þ ð6Þ Fig. 3 Disparity map for a single sub-volume

123
Exp Fluids (2008) 45:549–556 553

unfragmented disparity peak, especially when choosing a Choosing an initial volume larger than the illuminated
large number of sub-volumes. Figure 4 shows the volume is a good method to actually measure the size of
emerging disparity peak when adding disparity maps the illuminated volume or, respectively, the common area
from up to 16 images. The final outcome is one disparity viewed by all cameras. Figure 5 shows the disparity maps
map for each sub-volume in each z-plane for all cameras. for consecutive z-planes on one side of the volume indi-
Figure 5 shows the disparity maps for all sub-volumes for cating at what z-plane the illumination starts. The volume
one camera in a series of z-planes spaced 1 mm apart. The is also slightly tilted with respect to the coordinate system
off-center disparity peaks in the fully illuminated bottom as evidence by the left to right decrease in the disparity
z-plane correspond to a calibration error of up to 2 pixel. peak visibility.
Step f The sub-pixel disparity peak location is deter-
Step h The mapping functions Mi are corrected for each
mined, for example, by a 2D or dual 1D-Gauss fit and
camera separately according to Eq. 3. Since the disparity
corresponds to the most likely disparity vector di. The
vector fields are given only on grid positions defined by
disparity peak width is a good indication of the accuracy
the center of possibly very few sub-volumes one needs
of the 2D peak finder. For the experiment in Fig. 4, it is
to interpolate the vector field to intermediate positions.
only slightly larger than the selected Gauss width of 0.2
We choose to calculate new mapping functions M0 on the
pixel, suggesting that the 2D peak finder accuracy is
z-planes defined by the sub-volumes. For each z-plane an
around 0.1 pixel. In rare cases it has been observed that
array of 20 9 20 points is taken for which the difference
the disparity peak is smeared due to error gradients within
between the original mapping function and the disparity
a single (large) sub-volume. Usually, the calibration
field is computed and a new M0 is fitted using third-order
errors are highly uniform across the complete volume.
polynomial functions in x and y (similar to Soloff et al.
Step g The outcome is a 3D2C disparity vector field for
1997).
each camera, one vector for each sub-volume. This vector
Step i The sequence of steps a–h can be repeated after
field is validated using a 3D universal outlier median filter
the first self-calibration to check whether the disparity
(Westerweel and Scarano 2005) over a region of, for
peaks are finally positioned at the center position (0,0).
example, 3 9 3 9 3 or 5 9 5 9 5 vectors. One can store
One can choose in subsequent iterations a smaller
secondary disparity peaks in step f for possible selection
maximum allowed disparity emax and a lower threshold
by the median filter in case the highest peak is eliminated.
for more true particles and at the same time less ghost
If the selected volume is larger, for example, in z-direction
particles. Usually, two or three iterations are sufficient to
than the actual illumination, then complete z-planes of
reduce the mapping function errors down to below 0.1
sub-volumes might show random outlier vectors.
pixel. At the same time the number of sub-volumes and
The median filter with a kernel of 5 9 5 9 5 is robust the selected volume can be adjusted. Often it is
enough to throw out even complete z-planes of random advantageous to start with (e.g., 2 9 2 9 2) sub-
vectors. Note that a 3D median filter is considerably more volumes to first correct roughly the mapping functions
robust than a 2D filter due to the increased number of and use finer divisions (e.g., 10 9 10 9 10) later for a
neighbor vectors within some spatial domain. Most often it detailed local analysis of the 3D calibration errors.
is indicated to apply some additional 3 9 3 9 3 vector
smoothing to reduce the error inherent in the whole process
(Fig. 6). The true disparity field usually varies little across 3 Application to tomographic PIV experiments
the image. So it is possible to smooth the estimated dis-
parity field heavily to prevent errors propagating into an This procedure is tested for a tomographic PIV experiment
erroneous new mapping function. in water observing a transitional wake behind a finite

Fig. 4 Summing disparity maps over 1–16 recordings. Color scaling is always from minimum to maximum intensity

123
554 Exp Fluids (2008) 45:549–556

b Fig. 5 Disparity maps for decreasing depth z from top (not illumi-
nated) to bottom (fully illuminated). Sub-volume size in xyz is about
8 9 8 9 1 mm

length, circular cylinder with 3D instabilities at Re = 540


(Michaelis et al. 2006) with a volume of 73 9 46 9 16 mm
and a final 3D3C vector field with 141 9 89 9 29 vectors.
It turns out that the initial mapping function had errors of
up to 2–3 pixels close to the cylinder. After self-calibration
the errors are reduced to below 0.1–0.2 pixel.
The volume self-calibration procedure improves signif-
icantly the contrast of the reconstructed volume and the
quality of the vector field in terms of higher correlation
values and improved signal-to-noise ratios. the raw vector
field without any vector validation is shown in Fig. 7. Self-
calibration eliminates most outliers. The signal-to-noise
ratio of the (3D) vector correlation maps is almost twice as
large as before. Figure 8 shows the complete volume
vector field before and after the calibration procedure after
vector post-processing.
An important aspect is that without such a volumetric
self-calibration procedure one is not even aware of how
large the calibration errors actually are. As it turned out,
even with great care during the calibration procedure there
was hardly any experiment where the calibration errors did
not exceed 1 or 2 pixels in some corner of the volume.
In one experiment viewing with cameras from opposite
side of the volume it was possible to correct calibration
errors in the range of 7 pixels arising due to a double-sided
calibration plate tilted slightly relative to the z-direction in
which it was moved. This resulted in the two cameras on
the back side being shifted relative to the front side
cameras.
Not knowing how large the error actually was one
started out with emax = 3 pixel, not finding any peaks in the
disparity maps. Then one had to gradually increase emax
and carefully adjusts the particle detection threshold to
minimize the ghost particles and processing time. With
emax = 10 (400 times as many ghost particles as for emax =
3) and only using 2 9 2 9 2 sub-volumes it was finally
possible to see a faint disparity peak among the vast
amount of ghost particles. Once this large error had been
corrected in a first step it was then easy to further reduce
the calibration errors down to 0.1 pixels with fine local
resolution possible due to otherwise high-contrast, low-
noise images with particle densities less than 0.05 ppp.

4 Relevance for other PIV techniques

For standard stereo PIV and scanning light sheet stereo PIV
(Brücker 1996; Hori and Sakakibara 2004) the planar
stereo PIV self-calibration method is usually sufficient to

123
Exp Fluids (2008) 45:549–556 555

Fig. 6 Original disparity vector


field (left) and after vector
validation and smoothing (right)

measure and correct the position and rotation of the light self-calibration is then required to reduce the calibration
sheets which typically lead to disparities of 5–20 pixel. errors to acceptable levels.
Remaining calibration errors of the order of a few pixels Finally, volumetric self-calibration is also recommended
are only relevant in case of very strong local velocity for standard 3D-PTV. For low seeding densities and unique
gradients. But especially for thick light sheets required, for conversion of 2D particle positions into a 3D particle
example, for cross-flow measurements the volumetric self- position, a high systematic triangulation error is not criti-
calibration technique is recommended. Here standard pla- cal. But for higher seeding densities the correction of
nar self-calibration often fails if the light sheet profile is not calibration errors by volumetric self-calibration makes it
nicely Gaussian, but, for example, shows multiple intensity possible to choose a very small threshold for the maximum
peaks within the long correlation streak. allowed triangulation error. This significantly reduces the
Thick light sheet PIV images can be processed with the number of ghost particles and can provide some rejection
tomographic PIV algorithm usually adding a third or criteria for overlapping particles which otherwise lead to
fourth camera (Wieneke and Taylor 2006). Volumetric inaccurate 3D particle positions.
Table 1 summarizes the importance of the two self-
calibration algorithms for correcting laser sheet misalign-
ments and mapping function errors, respectively, for the

Fig. 7 Z-plane of volume vector field without (top) and with (bottom)
volume self-calibration. Shown is the raw vector field without vector Fig. 8 Complete 3D flow field without (top) and after self-calibration
validation with the 3D vorticity magnitude as background color and (bottom) with 3D vorticity colored according to dominant sense of
every fourth vector in x-direction rotation (Michaelis et al. 2006)

123
556 Exp Fluids (2008) 45:549–556

Table 1 Relevance of calibration errors for different techniques been reduced to 0.1–0.2 pixel. The method is also useful
Technique Inaccurate Misalignment
for stereo PIV especially with thicker light sheets since
mapping between here the standard planar self-calibration technique often
function calibration fails due to multiple intensity peaks within the light sheet.
plate and For 3D-PTV the method can reduce the number of ghost
light sheet
particles allowing higher seeding densities.
Stereo PIV - ++
Thick-sheet tomographic PIV ++ ++
Full-volume tomographic PIV ++ - References
3D-PTV + -
Brücker Ch (1996) 3-D scanning-particle-image-velocimetry: tech-
+ Strong, - Weak nique and application to a spherical cap wake flow. Exp Fluids
56:157–179
different planar and volumetric particle image velocimetry Calluaud D, David L (2004) Stereoscopic particle image velocimetry
techniques. measurements of a flow around a surface-mounted block. Exp
Fluids 36:33–61
Coudert S, Schon JP (2001) Back-projection algorithm with mis-
alignment corrections for 2D3C stereoscopic PIV. Meas Sci
5 Summary Technol 12:1371–1381
Elsinga GE, Scarano F, Wieneke B, van Oudheusden BW (2006)
Tomographic particle image velocimetry. Exp Fluids 41:933–
A volumetric self-calibration technique has been developed
947
to correct mapping function errors using the actual Hartley R, Sturm P (1994) Triangulation. ARPA Image Understand-
recordings of the particle images. This is especially rele- ing Workshop, Monterey, pp 957–966
vant for tomographic PIV where the mapping errors should Hori T, Sakakibara J (2004) High speed scanning stereoscopic PIV
for 3D vorticity measurement in liquids. Meas Sci Technol
be below 0.3 pixel. The method is based on detection of 2D
15:1067–1078
particle images and triangulation of the 3D particle position Maas HG (1996) Contributions of digital photogrammatry to 3D PTV.
in space. Particle images are usually intensity normalized In: Dracos T (ed) Three-dimensional velocity and vorticity
across the image and between cameras to account for measuring and image annalysis techniques. Kluwer, Dordrecht,
pp 191–208
Mie-scattering differences, and a threshold is used to select
Michaelis D, Poelma C, Scarano F, Westerweel J, Wieneke B (2006)
only the, for example, 10% brightest particles. Back-pro- A 3D time-resolved cylinder wake survey by Tomographic PIV.
jecting the 3D position toward the camera images and In: 12th international symposium on flow visualization, Göttin-
comparing to the original position provides a disparity gen, Germany, 10–14 September
Prasad A (2000) Stereoscopic particle image velocimetry. Exp Fluids
error for each particle. Collecting disparities for all parti-
29:103–116
cles throughout the volume and a suitable clustering Scarano F, David L, Bsibsi M, Calluaud D (2005) S-PIV comparative
method to eliminate false 3D particles one obtains a 3D2C assessment: image dewarping + misalignment correction and
disparity vector field for each camera. Sufficient statistics is pinhole + geometric back projection. Exp Fluids 39:257–266
Soloff SM, Adrian RJ, Liu ZC (1997) Distortion compensation for
obtained by summing the disparities from many recordings.
generalized stereoscopic particle image velocimetry. Meas Sci
After vector field validation and smoothing this disparity Technol 8:1441–1454
field is then used to correct the original mapping function. Westerweel J, Scarano F (2005) A universal detection criterion for the
It is also possible to measure the actual extent and rotation median test. Exp Fluids 39–6, 1096
Wieneke B (2005) Stereo-PIV using self-calibration on particle
of the illuminated volume from the disparity maps.
images. Exp Fluids 39:267–280
The technique has been applied successfully to a number Wieneke B, Taylor S (2006) Fat-sheet PIV with computation of full
of tomographic PIV experiments improving significantly 3D-strain tensor using tomographic reconstruction. In: 13th
the quality of the volumetric vector field. It was found that international symposium on applications of laser techniques to
fluid mechanics, Lisbon
in most experiments even with a careful calibration pro-
Willert C (1997) Stereoscopic digital particle image velocimetry for
cedure calibration errors of 1–2 pixel were found application in wind tunnel flows. Meas Sci Technol 8:1465–1469
somewhere in the volume. After self-calibration this has

123