Branley Coherer 1900: An Adventure in Photogrammetric Processing

Branley Coherer by NUIM_CallanMuseum on Sketchfab

This Branley Coherer was digitised as part of our AFF622 National Science Museum 3D recording project. The National Science Museum, affectionately known as the Callan Museum, is located on the South Campus Maynooth University, and curated by Doctor Niall McKeith. Ethan Fitzpatrick did the photographing of the Branley Coherer, producing 83 images which I then processed in Agisoft Photoscan.

This was the first processing of the Branley Coherer 1900. The settings used to generate the sparse cloud were 40,000 points and 4,000 tie points, with the setting set to high.The Dense Cloud was then generated at the high quality setting with aggressive depth fltering. From this dense cloud the mesh was generated, set to a high face count with interpolation enabled.  At this stage a number of issues presented themselves, namely the fact that a great deal of data was lost from the dense cloud particularly on the antennae, and the gap between the two black canisters on the right hand side of the object was filled. These issues led to the decision to re-process the object without texturing it.

The above images are from the second round of processing. This time, the sparse cloud was generated with the key point limit set to 100,000 points and the tie point limit set to 10,000 points. This was in the hope generating more data for the antennae, which was minimally effective. The dense cloud was generated with the same high and aggressive settings as the first time, and the mesh was generated with the same high face count. This time, however, mesh interpolation was disabled. This was an experiment which resulted in the gap between the two canisters being there, and was deemed successful. However, due to the insufficient data on the antennae it was felt that the object needed to be processed again.

This third round of processing, which involved the adjustment and refining of the original masks, resulted in an object which was minimally different from the second round of processing. Hence, the decision was taken to adjust the brightness and contrast of the images and re-process a fourth time.

In this final round of processing, I continued to use the alignment settings from the second round of processing, as well as the high aggressive dense cloud settings. For generating the mesh interpolation was again disabled, as it was found that doing so restored data to small parts of the mesh such as the glass tube between the antennae. This had become badly fragmented in the interpolated mesh, but maintained its integrity once the function was disabled. Editing the image did restore some data to the antennae, and improved the overall look of the object. However, there continued to be holes which we could not fill in Photoscan, and with interpolation disabled Autodesk Meshmixer was unable to cope with the object and crashed multiple times, due in part to the high number of points but also the number of issues which it identified in the mesh. Regardless of which settings were used to generate the mesh there existed a number of varied problems, with the antennae being an enduring issue. In the end it was felt that no one part of the object could be improved without compromising another part, and so the decision was taken to upload it to Sketchfab, with added annotations explaining the key issues.

And why, you may ask, did we have so many issues with the Branley Coherer? The answer comes back to data capture. The antennae were highly reflective, though at the time we thought we would be able to overcome this issue with image editing. If we were to re-capture the object we would use a polarising filter in order to minimise this reflectiveness. I suspect this would also have helped in maintaining the integrity of the glass tube.

Another issue is the simple composition of the object. There are many delicate parts and wires, and angles which it was difficult to capture, such as the inner sides of the canisters. Finally, there is a slight gap between the bell in the middle of the object and the wooden backing. This gap confused the software and led to it generating a hole when interpolation was disabled for the mesh. However, when interpolation was enabled the gap itself was filled and the integrity of the object remained compromised, albeit differently.

In conclusion, the Branley Coherer is an object which is difficult to accurately capture using photogrammetry. Its own structure complicates both capturing and processing, however I would not say the object was a complete failure. We did get a model of the Coherer, albeit one with a number of flaws, and in the process of doing so I learned a great deal not only about photogrammetry processing, but also about my patience for masking. This learning curve was one which proved useful when I worked on processing other models!

Particular thanks go to Doctor Niall McKeith for allowing us access to the Museum, Doctor Konstantinos Papadopoulos of An Foras Feasa, Maynooth University who is the Module Leaser for AFF622, and to Ethan Fitzpatrick who photographed the Branley Coherer.

Leave a Reply