GEOMETRIC CALIBRATION OF THE HASSELBLAD H3D

MEDIUM FORMAT CAMERA

A. Fauner, R. Ladstädter *, V. Kaufmann

Institute of Remote Sensing and Photogrammetry, Graz University of Technology, Austria

KEY WORDS: Photogrammetry, Digital, Camera, Calibration, Close Range, Aerial

ABSTRACT:

The Hasselblad H3D digital camera, which is designed for professional studio shootings, is equipped with a 39 megapixel digital
back. It would be desirable to use this very high geometric resolution in photogrammetric applications. Unfortunately, the digital
back does not have a tight connection to the camera body. In addition, the unknown internal image processing algorithms and the
autofocus lens make it difficult to obtain a reliable camera calibration. In this paper, we investigate whether the Hasselblad H3D
camera can be calibrated for use in photogrammetric applications. Repeated calibrations were performed using a 2D (planar) target
as well as a 3D test field. Results of the individual calibration projects are compared in order to analyze the stability of the camera
over time. We also investigate if there is a fixed pattern of systematic image residuals (after removal of radial distortion), which
could be described by a distortion grid. Software tools were developed for semi-automatic calibration using the MATLAB
programming environment. This calibration software consists of a graphical user interface supporting automated and precise
measurement of circular point targets, robust calculation of initial orientation parameters and bundle adjustment with self-calibration
capabilities. The commercial software package PhotoModeler (Eos Systems Inc./Vancouver) was also used for comparison to derive
calibration parameters based on the planar target.

1. INTRODUCTION

In this paper we present repeated geometric calibrations of a Raw image data was converted to RGB 8 bit TIFF for further
Hasselblad H3D digital SLR camera (see Fig. 1). The H3D photogrammetric processing. The final image size was 5412 x
digital back holds a 39 megapixel image sensor (36.7 x 49 mm) 7216 (file size of 111 MB for uncompressed storage).
and can be detached from the camera body for maintenance.
Our equipment includes three HC lenses, i.e., 3.5/35 mm, The H3D has been successfully used in terrestrial
3.5/50 mm, and 2.8/80 mm. Image data can be stored on photogrammetric projects, e.g. for glacier monitoring and
internal CF cards (type II) or on external storage media architectural projects, and in helicopter based mapping projects
(Imagebank or computer). Image data (16 bit, color) is stored in (Raggam et al., 2007). A relative accuracy of point positioning
lossless compressed Hasselblad 3F RAW file format. Automatic of at least 1:10000 is needed for such applications. A reliable
autofocus must be switched off for photogrammetric work. The camera calibration must thus be performed for all three
internal tilt sensor must also be disabled for proper available lenses. Furthermore, we wanted to investigate the
photogrammetric image orientation. Hasselblad FlexColor stability of the H3D camera by comparing repeated calibrations
software (version 4.6.7) is needed for further image processing. (in total four calibrations during 9 months).
Digital correction of the (lateral) effect of color aberration is
carried out by “Digital APO Correction” (DAC). In order to derive calibration parameters from images taken of
2D or 3D targets in reasonable time, it was necessary to
implement a (semi-)automatic workflow for point measurement,
calculation of approximate exterior orientation (EO) parameters
and final bundle adjustment. This was done in a master thesis
(Fauner, 2008) and a bachelor thesis (Längauer, 2008). Results
were cross checked by the commercial PhotoModeler software
package of Eos Systems Inc. (PhotoModeler, 2008).

Another aim of this study was to evaluate the quality of

calibration results using a low cost planar (2D) target and to
compare it with the results from a 3D test field. Special
investigations were carried out to evaluate systematic image
Digital back errors caused by the eccentricity error of the ellipse operator
(sensor unit)
Metal fixing bar
and distance dependant lens distortion.

Figure 1. Hasselblad H3D with fixing bar applied

* Corresponding author.

1077
 The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B1. Beijing 2008

2. CAMERA CALIBRATION markers are glued on aluminum bars mounted on a concrete

wall, floor and ceiling. The coordinates of all the targets have
been determined by a geodetic survey with a precision of +/-
2.1 Calibration using a movable planar target 0.05 mm. In contrast to the 2D calibration target, no coded
markers exist in the test field.
The calibration target is made of a stable planar wooden board,
sized 140 x 120 x 2.5 cm. Because of its low weight (<10 kg) it
can be easily moved to any location suitable for calibration,
indoor or outdoor. The layout of the present 2D target is based
on the target used by PhotoModeler. However, the square
layout of the original target was modified by adding two
additional columns of markers on the left and right side (see
Fig. 2). The four coded markers can be identified uniquely in
images taken from arbitrary viewing directions.

The rectangular pattern of 10 x 14 black markers (∅ 2cm)

allows acquiring single images fully covered by markers. This
facilitates measurement of the effect of lateral chromatic
aberration (cp. Kaufmann & Ladstädter, 2005).

Figure 3. Calibration room of Vexcel Imaging Graz

Images were taken from three positions (left, middle, right),

with different viewing directions. Again, additional images
were taken with the camera rotated 90°. Series of 15 to 35
images were produced for each lens (see Table 2). A high f-
number (f/27) was also used for the 3D test field. The limited
lighting required relatively long exposure times (~0.7sec) and
thus the use of a tripod. Because of the room dimensions, the
maximum distance to the back wall is limited to 8m.

2.3 Automated marker measurement

Figure 2. Layout of the 2D calibration target
Automated and precise measurement of circular markers
For calibration purposes, images should be taken from various requires an algorithm that is capable of fitting an ellipse in the
directions, viewing angles and distances. It is also image. Such an algorithm, described e.g. by Luhmann (1986),
recommended to take images with the camera rotated 90°. was implemented as a MATLAB function. Given a starting
Using such an image configuration, it is possible to de-correlate point somewhere inside the marker, the algorithm first
calibration parameters from EO parameters (e.g. focal determines the approximate (maximum) diameter of the ellipse,
length/object distance and principle point/rotation angles). The then performs a radial search for edges in various directions and
target should not be moved during calibration in order to avoid finally fits an ellipse to the previously located points (see Fig.
deformation. 4). Our test results show that the center of the ellipse can be
determined with an accuracy of at least one tenth of a pixel and
We performed three independent calibrations, on June 25, 2007, that the algorithm also works with low contrast or noisy images.
on July 1, 2007, and on March 19, 2008. Series of 16 to 31
images were taken handheld for each of the three lenses (see Fitted ellipse
Table 1). Using daylight, it was not necessary to use a tripod Center point
which speeded up the whole process. Images were taken using Approximation position
the far end (∞) of the focusing ring. This is a repeatable setting, Point measured on edge
Rays used for edge dedection
which corresponds more or less to an infinite object distance.
However, this setting causes blurred images because of the
close distance (max. 3m) of the calibration target, especially for
the 80mm lens. A high f-number (f/27) was therefore used to
minimize blurring.

2.2 Calibration using a 3D test field

Two separate calibrations were performed on August 1, 2007

and on March 19, 2008 (on the same day as for the second 2D
calibration) in the in-house calibration room of Vexcel Imaging
Graz (see Fig. 3), which is routinely used for calibration of the
UltraCamX digital aerial camera.. The size of the test field is Figure 4. Ellipse fitting operator
approximately 8m x 2.5m x 2.5m. A total of 394 circular

1078
 The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B1. Beijing 2008

It should be mentioned that the center of the ellipse does not This workflow was successfully implemented in a MATLAB
exactly coincide with the projected center of the circular tool (see Fig. 7). Image measurements for all images of the
marker. This small eccentricity is a function of the viewing calibration project are exported to an ASCII file for further
angle, marker size and focal length (Dold, 1997). The evaluation in the bundle adjustment.
evaluation of this formula (see Fig. 5) showed that maximum
eccentricity values of 1.25µm can be expected for the 2D target
(80mm lens) and even lower values (below 0.25µm) for the 3D
test field. The eccentricity error was therefore neglected for the
purposes of our study.
Eccentricity error
1

0.5
Eccentricity [μm]

-0.5
H3D 35 mm (Distance 4m)
H3D 50 mm (Distance 4m)
H3D 80 mm (Distance 4m)
-1
H3D 35 mm (Distance 2m)
H3D 50 mm (Distance 2m)
H3D 80 mm (Distance 2m) Figure 7. MATLAB tool for automated 2D measurements
-1.5
-80 -60 -40 -20 0 20 40 60 80
Viewing angle [deg] The following table gives an overview of the H3D calibration
Figure 5. Eccentricity error of the ellipse operator projects performed using the planar target:

Another error source concerning the ellipse operator was also Date Lens Focus # Images # Points
investigated. This error occurs when large radial distortion is June 25, 2007 35mm ∞ 16 2450
present in the images and the markers are imaged at a large 50mm ∞ 20 3150
image scale. In this case distortion may change even within a 80mm ∞ 19 2900
single imaged marker, which causes a non-linear deformation of July 1, 2007 35mm ∞ 17 2800
the ellipse, resulting in a positioning error. As can be seen from 50mm ∞ 18 3000
Fig. 6, this error is estimated to be well below 0.5µm for all 80mm ∞ 18 2800
lenses and can therefore also be neglected in this study. March 19, 2008 35mm ∞ 31 5200
50mm ∞ 23 3660
Distorted ellipse error 80mm ∞ 22 2950
0.5
35mm 1.5m 28 4660
Hasselblad H3D 35 mm
0.4
Hasselblad H3D 50 mm 50mm 2m 24 3800
0.3 Hasselblad H3D 80 mm 80mm 2m 19 2100
0.2
Table 1. Statistics of the 2D marker measurements
Positioning error [μm]

0.1

0
Automation of marker measurements is much more complicated
-0.1 for a 3D test field. We therefore decided to use a semi-
-0.2 automated approach where a limited number of manual
-0.3 measurements must be made for each image. Again, the
-0.4
necessary steps were implemented in a MATLAB tool (see Fig.
8):
-0.5
0 5 10 15 20 25 30
Radius [mm]
1. Manual measurement of at least four markers
2. Determination of approximate EO parameters using
Figure 6. Influence of large distortion on the ellipse operator the robust Müller/Killian resection algorithm
(modified after Killian, 1955: 171-179)
The following steps are necessary in order to fully automate the 3. Determination of refined EO and IO parameters and
process of marker measurement for a planar target (cp. radial distortion by a single image bundle adjustment
Längauer, 2008): (using additional manual measurements)
4. Re-projection of all visible markers into the image
1. Localization of (coded) markers in the image gives approximate positions for the ellipse operator
2. Determination of a 2D projective transformation 5. Automated precise marker measurement using the
between image and object coordinates ellipse operator
3. Estimation of radial distortion (optional)
4. Calculation of approximate positions for each marker Steps 3 to 5 can be performed iteratively until all markers have
5. Precise measurement using the ellipse operator been successfully measured. It is also possible to carry out steps
6. Determination of approximate EO parameters using a 1 to 4 for all images of the calibration project in advance and
robust resection algorithm (Killian, 1955: 97-104) run the (time consuming) automated marker measurement as a

1079
 The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B1. Beijing 2008

batch job. Image measurements, approximate EO parameters within the different projects. Pixel size is a physical
and marker coordinates are exported to ASCII files, which can constant, so we decided to fix this parameter at a
directly be used for bundle adjustment. mean value, resulting in a pixel size of 6.8 x
6.801µm.

3. The principal point offset is relatively large, which

causes problems when the principal point of
symmetry (PPS) is assumed to be at the image center.
We therefore set the PPS equal to PPA for all
calibration projects.

In a second and final run, all calibration projects were adjusted

using only the five significant parameters (and parameter c1
fixed at a given value). The remaining calibration parameters
are highly significant and much less correlated than the original
parameter set. The following tables present focal length and
principal point parameter values and sigma naught obtained
from the bundle adjustment:

June 25, 2007

Figure 8. MATLAB tool for automated 3D measurements Lens c [mm] x0 [µm] y0 [µm] σ0
35 mm 35.663 -50.1 279.6 0.9
The following table gives an overview of the H3D calibration ± 0.0007 ± 0.4 ± 0.4
projects using the Vexcel test field: 50 mm 50.286 -98.7 240.1 1.2
± 0.0011 ± 0.0006 ± 0.8
Date Lens Focus # # Points 80 mm 82.354 -176.7 302.3 2.9
Images ± 0.0056 ± 3.1 ± 4.6
August 1, 2007 35mm ∞ 15 4780 July 1, 2007
50mm ∞ 23 4900 Lens c [mm] x0 [µm] y0 [µm] σ0
80mm ∞ 16 1900 35 mm 35.642 -88.0 184.6 1.5
March 19, 2008 35mm ∞ 35 10600 ±.0012 ±0.6 ±0.7
50mm ∞ 20 4960 50 mm 50.280 -137.1 140.5 2.1
80mm ∞ 21 3100 ±0.0020 ±1.1 ±1.2
50mm 5m 18 4500 80 mm 82.272 -198.1 196.9 4.2
80mm 5m 23 3100 ± 0.0083 ±5.0 ±6.0
March 19, 2008
Table 2. Statistics of the 3D marker measurements Lens c [mm] x0 [µm] y0 [µm] σ0
35 mm 35.668 -70.4 294.9 0.7
± 0.0012 ± 0.5 ± 0.3
2.4 Determination of calibration parameters
50 mm 50.251 -105.8 233.3 0.7
The camera calibration parameters were determined using the ± 0.0008 ± 0.3 ± 0.5
in-house developed bundle adjustment software PhoBA 80 mm 82.292 -203.8 281.6 2.0
(Photogrammetric Bundle Adjustment). The following ± 0.0042 ± 2.0 ± 2.7
commonly used camera calibration parameters (cp. Luhmann et Table 3. Calibration parameters obtained from the 2D target
al., 2006) can be determined:

1. Focal length: c August 1, 2007

2. Principal point: x0, y0 Lens c [mm] x0 [µm] y0 [µm] σ0
3. Radial distortion: coefficients k0, k1, k2, k3 35 mm 35.652 -72.9 225.3 1.3
4. Tangential distortion: coefficients b1, b2 ± 0.0004 ± 0.2 ± 0.3
5. Affinity and shear: coefficients c1, c2 50 mm 50.251 -117.1 153.5 1.5
± 0.0004 ± 0.4 ± 0.5
In a first run, all of the calibration projects were adjusted using
80 mm 82.297 -174.2 270.5 1.4
the complete set of calibration parameters (except k0, which is
100% correlated with the focal length c). Gross errors were ± 0.0017 ± 1.3 ± 1.6
automatically removed by the built-in data snooping feature of March 19, 2008
PhoBA. From the analysis of these preliminary results, we Lens c [mm] x0 [µm] y0 [µm] σ0
could draw the following conclusions: 35 mm 35.651 -38.33 288.33 1.0
± 0.0002 ± 0.2 ± 0.2
1. Parameters b1, b2, k3 and c2 are not significant and can 50 mm 50.256 -85.99 225.85 1.3
be eliminated from the adjustment. ± 0.0006 ± 0.4 ± 0.4
80 mm 82.301 -132.47 314.26 1.1
2. Parameter c1 describing the deviation from a square
pixel (scale in x) is significant but varies slightly ± 0.0011 ± 1.1 ± 1.0
Table 4. Parameters obtained from the 3D test field

1080
 The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B1. Beijing 2008

Radial distortion coefficients k1, k2 are not given here target: 2m, 3D test field: 5m). Focal length parameters differ by
explicitly. Fig. 9 shows the unbalanced radial distortion curves about 0.4% for the 35mm lens, 0.7% for the 50mm lens and
derived from these parameters. 4.0% for the 80mm lens (see Table 5). With lenses focused at
Unbalanced radial distortion
5m, the differences are 0.2% and 1.4% (for the 50mm and
100
80mm lens, respectively, see Table 6).
0
PHOBA: Distance Dependent Radial Distortion
3.5
-100
Distance 1.25m
Distance 1.5m
-200 3
Distance 1.75m
Distance 2m
Distortion [μm]

-300 2.5

-400
2

dr [µm]
-500
Hasselblad H3D 35 mm (2D Calibration) 1.5
Hasselblad H3D 35 mm (3D Calibration)
-600
Hasselblad H3D 50 mm (2D Calibration)
Hasselblad H3D 50 mm (3D Calibration)
1
Hasselblad H3D 80 mm (2D Calibration)
-700
Hasselblad H3D 80 mm (3D Calibration)

0.5
-800
0 5 10 15 20 25 30 35
Radius [mm]

0
0 5 10 15 20 25 30
Figure 9. Unbalanced radial distortion curves Radius [mm]

A clear difference can be seen between the 2D and 3D Figure 11. Distance dependent distortion (2D)
calibration methods for all three lenses (calibration on March
19, 2008). This is caused by the high correlation of the radial Calibration results cannot be improved significantly and the
distortion parameters with other parameters in the bundle same systematic residuals can be observed in the images. For
adjustment (focal length, EO parameters), especially for the 2D the 80mm lens and the 2D target, however, results are much
target. Errors will also propagate into the object coordinates better, because we get sharp images with the focused lens.
(model deformation).
March 19, 2008
PHOBA: Single Image Residual Plot Lens c [mm] x0 [µm] y0 [µm] σ0
20 35 mm 35.820 -61.9 278.8 0.7
15
±0.0006 ±0.3 ±0.3
50 mm 50.585 -102.0 235.1 0.7
10 ±0.0007 ±0.3 ±0.4
5 80 mm 85.465 -243.0 256.5 0.9
±0.0028 ±1.1 ±1.6
0
y [mm]

-5 Table 5. Results with focused distance 2m (2D target)

-10
March 19, 2008
-15 Lens c [mm] x0 [µm] y0 [µm] σ0
-20 35 mm - - - -
50 mm 50.376 -86.3 229.9 1.2
-25
5µm ± 0.0006 ± 0.4 ± 0.4
-30
-30 -20 -10 0 10 20 30
80 mm 83.385 -112.9 310.1 1.0
x [mm] ± 0.0010 ± 0.8 ± 0.9
Figure 10. Systematic image residuals (3D target, 50mm lens)
Table 6. Results with focused distance 5m (3D target)
We observed small but systematic errors in some of the projects
when plotting image residuals for a single image (see Fig. 10).
We tried to model them by distance-dependent radial distortion 2.5 Quality check
parameters (cp. Dold, 2007) in the bundle adjustment.
In order to check the quality of the calibration process, all
As can be seen from Fig. 11, the effect of a distance-dependent calibration parameters are fixed at their estimated values in the
variation of radial distortion is not very significant. Maximum bundle adjustment. The 3D test field is now used to check both
values of 3µm (for the 2D target) are reached, but only in the the 2D and 3D calibration results (of March 19). Only nine
very corners of the image. The effect is even smaller for the 3D (well distributed) control points are used, the other 383 points
target (2µm). are introduced as check points to evaluate the accuracy of point
reconstruction (see Tables 7 and 8). Only three images are used,
In an additional experiment, we performed complete i.e., one for the left, middle and right position, which can be
calibrations with lenses focused at a mean object distance (2D seen as a “common” (not highly redundant) stereo
configuration.

1081
 The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B1. Beijing 2008

Lens RMS_X RMS_Y RMS_Z σ0 criterion is not met for the 2D calibration, whereas relative
[mm] [mm] [mm] accuracies of up to 1:20000 (planar positioning) and 1:4000 (Z
35 mm ±1.95 ±1.08 ±1.90 1.8 component) are achieved using the 3D calibration.
50 mm ±0.62 ±0.34 ±0.92 1.6
80 mm ±0.51 ±0.62 ±0.80 0.9 The H3D camera can therefore be used for close range
applications (e.g. architectural photogrammetry) without self-
calibration. If parameters of a 2D calibration are used, self-
Table 7. Quality check of the 2D camera calibration
calibration (focal length and radial distortion parameters)
should be applied if reliable control points are available. This is
Lens RMS_X RMS_Y RMS_Z σ0
also recommended for aerial (small scale) projects, even for 3D
[mm] [mm] [mm]
calibration.
35 mm ±0.30 ±0.18 ±0.64 0.5
50 mm ±0.34 ±0.32 ±0.64 0.6
80 mm ±0.19 ±0.34 ±0.59 0.5 4. REFERENCES

Table 8. Quality check of the 3D camera calibration Dold, J., 1997: Ein hybrides photogrammetrisches Industrie-
messsystem höchster Genauigkeit und seine Überprüfung.
Schriftenreihe, Studiengang Vermessungswesen, Universität der
3. CONCLUSIONS Bundeswehr München, Heft 54, pp. 26-32.

Marker measurement can be fully automated using the planar Fauner, A., 2008: Erstellung einer Applikation zur Kalibrierung
2D target. PhotoModeler and our in-house developed software von Digitalkameras. Unpublished master thesis, Graz
give very similar results. In the 3D case, automation is more University of Technology, in German, 91 p.
difficult and error prone. We will need to further improve our
software in order to reduce the number of gross errors (up to Kaufmann, V., Ladstädter, R., 2005: Elimination of color
5%). fringes in digital photographs caused by lateral chromatic
aberration. Proceedings of the XX International Symposium
If the ellipse operator is used for automated point CIPA 2005, 26 September - 1 October 2005, Turin, Italy, Vol.
measurements, the marker size must be chosen carefully, 1, pp. 403-408.
depending on the image scales used. If the size of the imaged
markers gets too small, the accuracy of the measurement will Killian, K., 1955: Über das Rückwärtsschneiden im Raum.
decrease. If markers get larger, systematic errors will increase Österreichische Zeitschrift für Vermessungswesen, 43, pp. 97-
(eccentricity error and ellipse deformation caused by radial 104, pp. 171-179.
distortion). A marker diameter of 20-50 image pixels has been
found to be the optimal size. Längauer, K., 2008: Entwicklung von MATLAB-Routinen zur
automatischen Bildkoordinatenmessung von kreisförmigen
Focal length and radial distortion are stable and can be Mustern in Digitalbildern für Zwecke der Kamerakalibrierung.
determined very accurately using the 3D test field. Using the Unpublished bachelor thesis, Graz University of Technology, in
2D target, these parameters show a much higher variation and German, 39 p.
lower significance. The principal point is very unstable, varying
up to 50µm between calibrations. This has been expected Luhmann, T., 1986: Ein Verfahren zur rotationsinvarianten
because of the unstable connection between the camera body Punktbestimmung. Bildmessung und Luftbildwesen, 4/1986, pp.
and the digital back. It is therefore necessary to recalibrate the 147-154.
camera each time the digital back has been removed (e.g. for
sensor cleaning). Luhmann, T., Robson, S., Kyle, S., Harley, I., 2006: Close
Range Photogrammetry: Principles, Methods and Applications.
A significant scale difference (1.5E-4) has been determined for Whittles Publishing, Dunbeath, Scotland, UK.
the x and y component of the sensor. This scale difference adds
up to 7.2µm (more than one pixel) for the longer sensor PhotoModeler, 2008: http://www.photomodeler.com (accessed
dimension. This results in a non-square pixel size of 6.800 x 1 April 2008)
6.801µm.
Raggam, H., Wack, R., Gutjahr, K., 2007: Mapping Capability
Setting focus to infinity for calibration causes problems when of a Low-Cost Aerial Data Acquisition Platform - First Results.
close range targets are used. Images out of focus will be ISPRS-Workshop, Commission VI, WG VI/4, "High Resolution
blurred, especially when narrow angle lenses are used and Earth Imaging for Geospatial Information", Hannover, May
object distance gets too short. Although the ellipse fitting 29th - June 1st, 2007.
operator can handle blurred images to a certain extent, accuracy
of image measurements will be reduced (e.g. for the 80mm lens
using the 2D target). 5. ACKNOWLEDGEMENTS

The quality of the calibration can be checked independently

only for the 2D calibration (in the 3D case, the same images are We wish to thank Dr. Michael Gruber from Vexcel Imaging
used for calibration and quality check). A relative accuracy of Graz (a 100% subsidiary of Microsoft Corporation) for the
1:10000 equals 0.8/0.5/0.4mm (for the 35/50/80mm lenses, opportunity provided to use their in-house calibration test field.
respectively) in object space (planar component) and 0.25mm (z
component). As can be seen from Tables 7 and 8 this quality

1082