3D Digitisation of the Archaeological Museum of Milan

View of the archaeological museum of Milan, inner courtyard

The research group of Politecnico di Milano   (Gabriele Guidi, Sara Gonizzi, Laura Micoli, Michele Russo and Davide Angheleddu) is digitising the Archaeological Museum of Milan and its collections in 3D.  The museum is housed in buildings of the monastery of San Maurizio which date from the 16th to 18th centuries and were built on the remains of a 4th century Roman circus. Remains of Roman houses dating to the 1st century are also visible in the cloister of the museum.

The museum holds more than 500 archaeological objects from the Roman, Greek, Etruscan, Barbaric and Medieval periods).  The collections include mosaics, furniture, ceramics, statues and building fragments - even statues which were later reused as building blocks!

The survey

The Archaeological Museum is organized in thematic rooms, one for each historical period, with the Roman age covering the ground floor and the two courtyards, the Medieval period on the first floor, Etruscan on the second floor and the Greek period on the top floor.  Some objects are fixed on pillars or on the walls, others are movable or in cases. Rooms are lit with spotlights pointed directly on the objects - on the ground floor there is a big glass wall forming one side of the room.

During the survey, these logistics strongly influenced the acquisition of the images:

  • it was impossible to cover the entire surface of objects which were fixed against walls resulting in the presence of holes and gaps in the final mesh;
  • the illumination gave other problems, due to the changes of the colour of the object itself.
Sculpted head of Jupiter positioned close to a column

Sculpted head of Jupiter positioned close to a column

Rear view of the mesh showing the gap caused by lack of data due to the occlusion

Rear view of a Sculpture head showing the gap in the model caused by the original being fixed close to a wall

Not much can be done to produce a complete model where objects are fixed against walls.  This is not too problematic with flat objects such as stele or inscriptions, which have nothing behind except a rough surface that can be easily replaced in the model.  It is obviously more problematic with sculptures where the object is all-round.  In cases such as the head illustrated here, the model can be closed but of course is somewhat false in respect to the original sculpture.

The fixed position  of some objects imposed also a short camera-object distance, and therefore the choice of the right type of camera and a lens short enough to frame a suitable portion of the object, was a critical point.

In this project three digital reflex and a mirrorless camera has been used: a 5D Mark II, a 20D and a 60D all by Canon, and a NEX5 by Sony, with lenses ranging from 18 to 50 mm, and different sensors formats (full frame for the 5D and APS-C for the others). It is important to notice that with APS-C the area covered is definitely smaller than the full frame, and maintaining the same lens the angle of view is proportionally smaller. This fact involves a multiplying factor of 1.6 for evaluating the equivalent focal length on a full frame camera. In other words a 20mm lens used on a 60D covers the same image area of a 32mm on a 5D. For this reason the most critical unmovable artefacts have been imaged with 5D+20 mm or Nex5 + 18 mm, providing a viewing angle of 84° and 66° respectively.

The changes in lighting between two images of the same objects acquired from different points of view

Illumination   in the museum also influenced the survey, especially when the spotlights were directly oriented on objects. In this case the illuminating conditions may prevent the automatic algorithm to identify corresponding portions of the object. A big problem was the possible presence of windows that created a significant backlight effect compromising sometimes the shots. The solution in those cases was to use a flat panel to shield the backlight avoiding the strong light imbalance and the relatively dark foreground.

Data processing

The Agisoft Photoscan software was used for data processing. This is a semi-automatic software in which both the external alignment of the camera and the internal calibration are done with few interaction by the user. Some choices can be done during image orientation, where the operator may set: i) alignment accuracy level; ii) possible control points; iii) hide image portions to mask the (sometime misleading) subject surrounding area. At mesh generation stage, the software permits to decide the accuracy and the polygon number of the final 3D model. The software implements image orientation and mesh generation through SFM and dense multi-view stereo-matching algorithms (Exact, Smooth, Height Field and Fast).

Creation of models

Masks were used for the entire process thus avoiding any bad alignments that would have been caused by the software taking points from the surroundings, or poor quality meshes.

The problems which occurred during the survey stage also conditioned also the processing of the data. The selection of the camera and the appropriate lens was easily performed after few tests on the same objects acquired with different settings and also on different items, to understand the best set up for different shapes.

As a software derived from the SFM philosophy, Agisoft works better with many images taken with a short baseline  rather then few images with a relatively long baseline as in standard photogrammetry. This involves a significant overlapping  among images that eases the automatic image matching, preferably on more than two images. In situations where the narrow viewing angle and the environmental constraints imposed strong limitations in the shooting position, the software had problems in finding the homologous points and the alignment was not satisfactory, preventing the following mesh generation.

Comparison between models produced in bad illumination (left) and using lighting panels and the right paramater settings for creating the mesh in Agisoft (right)

Comparison between models produced in bad illumination (left) and using lighting panels and the right paramater settings for creating the mesh in Agisoft (right)

About the illumination contrasts - it was decided to use shielding panels  to avoiding backlights and reflecting panels  to make more homogeneous the object lightening by brightening the darker shades. During the 3D processing it was found that the excessively dark shadows on the objects influenced the results creating a rough surface (left image). Where a better lightning was not available, the only possibility for obtaining a better result was to use the “smooth” geometry processing (right image).

The 3D Icons pipeline

Several tests were done using items different in shape, position and size, to test the software potential with this kind of objects and the best pipeline for producing a huge amount of accurate model in the shortest possible time frame. The first step was a suitable image masking  to reduce both the number of pixel processed (i.e. the workload) and the possible interference of surrounding elements in the scene on the main subject.

Mask drawn around the main subject

Mask drawn around the main subject

The masked images can be used for the first processing step, calculating automatically the alignment  of several images taken around the object to be modelled, based on matching features between adjacent shots.

 The following step consist in a dense image matching  that allows to calculate coloured 3D points by triangulating different point of views, possibly followed by the generation of a triangular mesh from the 3D point cloud (Build Geometry). This phase can be parameterized depending on the actual shape to be generated:

  • “Object type” can be set to “Arbitrary” (a 3D free form) or “Height Field” (a 2.5D surface like a DTM);
  • “Geometry type” can be set as “sharp” or “smooth” depending on the shape of the object to be modelled;
  • “Target quality” specifies the desired mesh quality: higher quality settings can be used to obtain more detailed and accurate geometry, but require longer time for processing
  • “Face count” specifies the maximum number of polygons in the final mesh and 0 indicates that no decimation is set
  • “Filter threshold” specifies the maximum face count of small connected components to be removed after surface reconstruction (as percentage of the total face count). The 0 value disables connected component filtering.

In the specific 3DIcons pipeline for the museum objects, after images orientation, the mesh generation was made with “Object type”=Arbitrary; “Geometry type”=sharp; “Target quality”=Medium; “Face count”=0; Filter threshold”=0. The only variable parameter was the “Geometry type” that was established to be changed from “sharp” to “smooth” according to the type of object to be modelled.

With these parameters, the results were high polygons meshes with a good accuracy. After the processing with Agisoft, the models were saved with image texture (arranging the image as 4096x4096 pixel) in obj format. The result was then imported in Polyworks to correct possible topological errors and to close gaps and lacking data omitted due to the environmental constraints. Finally a polygon decimation was made for avoiding excessive polygon densities for flat or smooth geometries as naturally generated by the image matching stage. The latter step was necessary because the aim of the 3D Icons project is to create light and manageable 3D models to be put on the internet and be easily opened on average personal computers.

Final model of an Altar from the collections at Milan archaeological museum

Final model of an Altar from the collections at Milan archaeological museum

Tomb decoration

Tomb decoration

Decoration of a public building

Decoration of a public building

Sculpture of Hercules

Sculpture of Hercules

Head of Massimiano

Head of Massimiano

Head of Jupiter

Head of Jupiter

Example of models generated through the optimized SFM process described above

Acknowledgements

The authors want to thank the director of the Archaeological Museum of Milan Dr. Donatella Caporusso for authorizing the surveys, all the staff at the Museum and Dr. Laura Micoli for her support in testing Agisoft and during the surveys at the Museum.