Image-based reconstruction ofthe Great Buddha of Bamiyan, Afghanistan

Armin Gruen, Fabio Remondino, Li Zhang

Institute of Geodesy and PhotogrammetrySwiss Federal Institute of Technology Zurich

ETH Hoenggerberg8093 Zurich, Switzerland

E-mail: <agruen><fabio><zhangl>@geod.baug.ethz.chWeb: http://www.photogrammetry.ethz.ch/

ABSTRACTIn the great valley of Bamiyan, north-west of Kabul, Afghanistan, two big standing Buddha statues were carved out ofthe sedimentary rock of the region around the second to fourth centuries AD. The larger statue was 53 meters high whilethe smaller Buddha measured 35 m. The two colossal statues were demolished on March 2001 by the Taleban, usingmortars, dynamite, anti-aircraft weapons and rockets. After the destruction, a consortium was founded to rebuild theGreat Buddha at original shape, size and place. Our group performed the required computer reconstruction, which servesas a basis for the physical reconstruction. The work has been done using three different types of imagery in parallel andin this paper we present our results of the 3D computer reconstruction of the statue.

Key words: Calibration, Orientation, Surface Reconstruction, Modeling, Cultural Heritage

1. INTRODUCTIONThe town of Bamiyan, 200 km north-west of Kabul, was one of the major Buddhist centres from the 2nd century AD upto the time when Islam entered the valley in the 9th century. The town was situated in the middle of the Silk Route and itwas a common meeting place for many ancient cultures and later tourists. In the north valley of Bamiyan, two big standingBuddha statues were carved out of the sedimentary rock of the region, at ca 2500 meters of altitude. The Emperor Kan-ishka ordered their construction around the 2nd century AD. Some descendants of Greek artists, who went to Afghanistanwith Alexander the Great, started the construction that lasted till the 4th century. The larger statue was 53 metres highwhile the smaller Buddha measured 35 m. They were cut from the sandstone cliffs and they were covered with a mixtureof mud and straw to model the expression of the face, the hands and the folds of the robe. The two giants were painted ingold and other colours and they were decorated with dazzling ornaments. They are considered the first series of colossalcult images in Buddhist art.The two statues were demolished on March 2001 by the Taleban, using mortars, dynamite, anti-aircraft weapons and rock-ets. The Buddhists, the world community, ONU and UNESCO failed to convince the Taleban to leave such works of cul-tural heritage and now they are nothing but piles of sandstone rubble and rocks.After the destruction, a consortium was established with the goal to rebuild the Great Buddha of Bamiyan at originalshape, size and place. This initiative is lead by the global heritage Internet society New7Wonders [7], with its founderBernard Weber and the Afghanistan Institute & Museum, Bubendorf (Switzerland), with its director Paul Bucherer. Ourgroup has volunteered to perform the required computer reconstruction, which will serve as a basis for the physical recon-struction. We have produced various 3D models of different quality, using different types of images and based on auto-mated or manual image measuring techniques. Using our model generated with manual measurements on metric images,first a statue at 1/10 of the original size will be built and displayed in the Afghanistan Museum in Switzerland. This willbe used to study materials and construction techniques to be applied in the final rebuilding at full size in Bamiyan. In thispaper we present our results of the computer reconstruction of the Great Buddha in Bamiyan.

Figure 1: Left: The Great Buddha of Bamiyan before the destruction. Center: The explosion of March 2001. Right: The hole left in thecliff after the destruction.

2. IMAGES OF THE GREAT BUDDHA OF BAMIYANOur work is based on the use of three different types of imagery in parallel:1. A set of images acquired from the Internet (“Internet images”);2. A set of tourist-type images acquired by Harald Baumgartner, who visited the valley of Bamiyan between 1965 and1969 (“Tourist images”);3. Three metric images acquired in 1970 by Prof. Kostka, Technical University of Graz [6].The computer reconstruction is performed with automatic and manual procedures. We are still processing the tourist im-ages, while results from the other data sets are already available. Only the computer model generated with manual meas-urements on the metric images will be used for the physical reconstruction of the statue, but in this paper we report all theresults obtained on the different available image data sets.Originally our interest in the reconstruction of the statue was a purely scientific one. We planned to investigate how suchan object could be reconstructed fully automatically using just amateur images taken from the Internet. Then, after learn-ing about the efforts to actually rebuild the Great Buddha we decided to get involved in the project beyond a purely sci-entific approach and to contribute as much as we could with our technology to the success of the work.

3. PHOTOGRAMMETRIC PROCESSINGThe reconstruction process consists of phototriangulation (calibration, orientation and bundle adjustment), image coordi-nate measurement (automatic matching or manual procedure), point cloud and surface generation, texture mapping andvisualization. From [6], a contour plot of the statue (scale 1:100) was available. From this plot we derived some controlpoints used in the orientation procedure.

3.1. The Internet data set

The main scientific challenge here lies in the facts that no typical photogrammetric information about these images isavailable and that existing automated image analysis techniques will most probably fail under the given circumstances.Out of 15 images found on the Internet, four were selected for the processing (Figure 2). Two in front of the Buddha, onefrom the left side and one from the right side. All others were not suitable for photogrammetric processing because of verylow image quality, occlusions or small image scale.

Figure 2: The Internet images used for the 3D reconstruction of the statue.

The main problems of these images are their differences in size and scale, the unknown original format, pixel size andinterior orientation and most of all the different times of acquisition. Therefore some parts visible in one image are missingin others. Also, the illumination conditions are very different and this can create problems with the automatic matchingprocedures.

3.1.1. Calibration and Orientation

For the interior orientation, pixel size and focal length of each image were assumed, as well as the principal point, fixedin the middle of the images. With this last assumption, we considered the size of the found images as the original dimen-sion of the photo, while they could be just a part of an originally larger image. The assumed pixel sizes are between 0.03mm and 0.05 mm.Concerning the exterior orientation parameters, as no information was available, we first performed an interactive deter-mination of the camera positions, varying also the value of the focal length and using some control points measured onthe contour plot of Prof. Kostka. Then we refined these approximations with single photo spatial resection solutions.The final orientation of the Internet images was achieved using bundle adjustments while the required tie points betweenthe images were measured semi-automatically with an Adaptive Least Squares matching algorithm. The final averagestandard deviations of the object point coordinates located on the Buddha statue and on its immediate vicinity was σx=0.08m, σy=0.17 m, σz=0.3 m. These are absolute precision values, which are however not representative for the reconstructionaccuracy of the statue. The latter must be based on relative precision considerations and is as such much better. The re-stored configuration of the cameras is shown in Figure 3.

Figure 3: Two views of the restored camera poses of the Internet images.

3.1.2. Surface Reconstruction

A Multi-image Geometrically Constrained Least Squares matching [1, 3] software package developed at our Institute wasapplied to the Internet images. The automatic surface reconstruction works in fully automated mode according to the fol-lowing procedure [4]:1. Selection of one image as the master image2. Extraction of a very dense pattern of feature points in the master image using the Moravec operator. At this stage, the

master image is subdivided into 7x7 pixel image patches and within each patch only one feature point which gains thehighest interest value is selected.

3. For each feature point, using the epipolar geometry determined in photo-triangulation, we get the approximatematches for the following MPGC (Multi-Photo Geometrically Constrained) matching procedure by a standardcross-correlation technique (Figure 4).

4. MPGC is applied for fine matching, including patch reshaping (Figure 4). MPGC exploits a priori known geometricinformation to constrain the solution and provides for the simultaneous use of more than two images [1, 3].

The difficulties of this data set lie in the large differences between the images, due to the different acquisition time, thechanging illumination conditions and the different image scales.

Figure 4: Upper part: Epipolar geometry between the images to get the approximate values for the matching.Lower part: MPGC matching results and computed 3D coordinates via forward intersection.

Three images are used for the matching procedure and a point cloud of ca 6000 points is obtained (Figure 5, left). Someholes are present in the results because of surface changes (due to the different time of image acquisition) or to the lowtexture in some areas. For the conversion of the point cloud to a triangular surface mesh, a 2.5D Delauney triangulationis applied (Figure 5, central). The final texturized 3D model is shown in Figure 5, right.

Figure 5: Left: Point cloud obtained from the Internet images (ca 6000 points). Central: The triangulated points visualized in wireframemode. Right: The texturized 3D model of the Great Buddha obtained from the Internet images.

3.2. The metric data set

The metric images were acquired in August 1970 with a TAF camera [2] by Prof. Kostka (Technical University of Graz).The TAF (Terrestrische Ausrüstung Finsterwalder) is an old-fashioned photo-theodolit camera that acquires photos on13x18 cm glass plates. The original photos were scanned by Vexcel Imaging Inc with the ULTRA SCAN 5000 at a reso-lution of 10 micron. The resulting digitized images resulted in a format of 16930 x 12700 pixels each (Figure 6).

Figure 6: The three metric images (A, B, C) acquired by Prof.Kostka in 1970.

3.2.1. Calibration and Orientation

Concerning the interior orientation parameters, the focal length and the principal point of the TAF camera were known,as well as the acquisition procedure [5, 6]. The images were acquired in normal case, with a double baseline and a distanceof ca 130-150 m from the statue. Using this information and some control points measured on the contour plot we achievedthe first approximations of the exterior orientation. The final orientation of the images was achieved using a bundle ad-justment with 35 tie points and six control points. The final standard deviation a posteriori of unit weight was 0.019 mmwhile the average standard deviations of the object point coordinates located on the Buddha itself and on its immediatevicinity was σx,y= 0.07 m and σz= 0.14 m. Again these values are not representative for the reconstruction accuracy, whichshould be based on relative precision values, rather than on absolute ones.

3.2.2. Surface Reconstruction with automatic matching procedure

The 3D model of the Buddha statue was automatically generated with the digital photogrammetric system VirtuoZo. Thematching method used by VirtuoZo is a global image matching technique based on a relaxation algorithm [9]. It uses bothgrid point matching and feature point matching.The program first uses a feature point based matching method to compute a relative orientation between couples of imag-es. Then the measured features are used to weight the smoothness constraints and utilized as approximations in the fol-lowing global matching method [8]. In our application, the images B and C of the metric data set were used forreconstruction. A regular image grid with 9 pixels spacing was matched using a patch size of 9x9 pixels and four pyramidlevels.The software has a pre-processing module which allows to apply a smoothness constraint satisfaction procedure. With thissmoothness constraint, poor texture areas can be bridged, assuming that the model surface varies smoothly over the imagearea. Through the VirtuoZo pre-processing module, the user can manually or semi-automatically measure some featureslike ridges, edges and regions in difficult or hidden areas. These features are then used as breaklines and planar surfacescan be interpolated (e.g. between two edges).As a result we obtained a very dense point cloud of ca 178 000 points, modeling the entire Buddha statue (however, notin full completion), including parts of the surrounding rocks. The statue as well as the rock around it are quite well recon-structed. But, due to the smoothness constraint and grid-point based matching, the small folds on the body of the dress andother very small features were not rendered. Therefore these important small features had to be measured manually.For the modeling, a 2.5D Delauney triangulation is performed. For that the 3D surface model is expanded to a plane bytransforming the point cloud from the cartesian coordinate system onto a cylinder coordinate frame. In the defined ρθζcylinder frame, ζ is the vertical cylinder axis crossing the model center and parallel to the original Y-axis of the cartesianobject coordinate system. ρ is the euclidean distance from the surface point to the z-axis and θ is the angle around theζ-axis. The 2.5D Delauney triangulation was done in the θζ plane and the final shaded model of the triangulated mesh isshown in Figure 7 (upper line, right). Then the central image of the metric data set is mapped onto the 3D geometric sur-

face to achieve a photorealistic virtual model (Figure 7, lower part). The lower part of the legs is not modeled because inthe used stereomodel the legs were not completely visible.

Figure 7: Upper line, left: Point cloud obtained with an automatic matching procedure (ca 178 000 points).Upper line, right: The shaded model generated after the triangulation of the points.

Lower part: The texturized 3D model of the Great Buddha obtained from the metric images.

3.2.3. Surface Reconstruction with manual stereo measurements

As mentioned before, many important small features could not be reconstructed by automated area-based matching. Forexample, the dress of the Buddha was rich in folds, which were between 5 and 10 cm in width (Figure 8).

Figure 8: Left and Central: The folds on the dress of the Buddha. Right: The manual measurement of the points on the statue.

Only accurate manual measurements could precisely and reliably reconstruct the shape and curvature of the dress. There-fore the metric images were imported to the VirtuoZo stereo digitize module and manual stereoscopic measurements wereperformed (Figure 8, right).The three stereo-models (A/C, A/B and B/C) were set up and points were measured along horizontal profiles of 20 cmincrement while the folds and other main edges were measured as breaklines. With the manual measurement a point cloudof ca 76 000 points was obtained (Figure 12) and the folds on the dress are now well visible. In the following surface tri-angulation, the 3D shapes of the folds were completely reconstructed and are visualized in wireframe format in Figure 12.The final surface model is presented in Figure 10, in smooth shaded and texturized mode.These results of the manual measurements of the metric images will be used for the physical reconstruction of the statue.

Figure 9: Point cloud of the manual measurements and triangulated surface displaying the reconstructed folds of the dress.

Figure 10: Left: Shaded model of the reconstructed statue with part of the surrounding rock.Center: Shaded model of the full statue, including part of the ground.

Right: Texturized 3D model of the Great Buddha of Bamiyan reconstructed from the metric images with manual measurements.

4. CONCLUSIONSThe photogrammetric reconstruction of the Great Buddha of Bamiyan has been achieved successfully. We have producedvarious 3D computer models of different quality, based on automated or manual image measuring techniques and usingdifferent types of images. The automated point cloud generation methods (with the Internet and the metric images) providefor dense point clouds (in presence of good texture) but they can fail to extract the very fine details of the statue, like thefolds of the robe. Moreover, also other important edges are often missed or smoothed. Therefore only manual measure-ments allow the generation of 3D models which are precise, reliable and complete enough to serve as the basis for a phys-ical reconstruction. With a pixel size of 10 micron (ca 1 cm on the object) manual measurements can be done with arelative accuracy of about 1-2 cm. While such high accuracy is necessary to model the folds of the robe (5 - 10 cm inwidth) correctly, it is surely more than sufficient to represent the overall form of the 53 m high statue in very close resem-blance to the original.The problems encountered here with the orientation of amateur images and with automated matching could be solved inan acceptable manner. The main difficulties of this project consisted in the transition from the point cloud (includingbreaklines) to a surface model which can satisfy high modeling and visualization demands.A web site of the work has been established on our server and is available at:http://www.photogrammetry.ethz.ch/research/bamiyan/ including animations, video, 3D models and other details.Following our computer reconstruction, the second phase is concerned with the construction of a 1/10 scale model, whichwill be set up in the courtyard of the Afghanistan Museum in Bubendorf, Switzerland. The 1/10 scale statue will be usedto study the materials and the construction methods to be applied in the final reconstruction in Bamiyan and for the im-plementation of the necessary infrastructures. In the meantime, UNESCO has moved forward to support the reconstruc-tion, which will hopefully lead to a soon initialization of the physical reconstruction.

REFERENCES

1. Baltsavias, E., Multiphoto Geometrically Constrained Matching, Dissertation, IGP, ETH Zürich, Mitteilungen No. 49,221 pp, 1991

2. Finsterwalder, S., Hofmann, W., Photogrammetrie. De Gruyter Lehrbuch, Berlin, pp. 119-120, 19683. Grün, A., Baltsavias, E., Geometrically Constrained Multiphoto Matching, Photogrammetric Engineering and Remote

Sensing, Vol. 54, No. 5, pp. 633-641, 19884. Grün, A., Zhang, L., Visnovcova, J., Automatic Reconstruction and Visualization of a Complex Buddha Tower of Bay-

on, Angkor, Cambodia, Proceedings of 21th Wissenschaftlich Technische Jahrestagung of Deutsche Gesellschaft fürPhotogrammetrie und Fernerkundung (DGPF), 4-7 September, Konstanz, Germany, pp.289-301, 2001

5. Grün, A., Remondino, F., Zhang, L., Reconstruction of the Great Buddha of Bamiyan, Afghanistan, International Ar-chives of Photogrammetry and Remote Sensing, 34(5), pp. 363-368, Corfu (Greece), 2002

6. Kostka, R., Die stereophotogrammetrische Aufnahme des Grossen Buddha in Bamiyan, Afghanistan Journal, Vol.3,nr.1, pp. 65-74, 1974

7. New7Wonders Foundation: http://www.new7wonders.com8. VirtuoZo NT, 1999, Version 3.1 Manual, Supresoft Inc.9. Zhang, Z., Zhang, J., Wu, X., Zhang, H., Global Image Matching with Relaxation Method, Proceedings of the Interna-

tional Colloquium on Photogrammetry, Remote Sensing and Geographic Information Systems, 11-14 May, Wuhan,China, pp. 175-188, 1992