Affordable digitisation of insect collections using photogrammetry

Insects dominate animal biodiversity and are sometimes called "the little things that run the world". They play a disproportionate role in ecosystem functioning, are highly sensitive to environmental change and often considered to be early indicators of responses in other taxa. There is widespread concern about global insect declines^[1] yet the evidence behind such declines is highly biassed towards the Global North and much is drawn from short-term biodiversity datasets^{[2] [3]}.

The Insect Collection at the University Museum of Zoology, Cambridge holds over 1.2 million specimens. These include specimens collected from the early 19th century to the present day. Most specimens remain undocumented and unavailable for analysis. However, they contain data that are critical to understanding long-term species and community responses to anthropogenic change, and vital to evaluating whether short-term declines are representative of longer-term trends^{[4] [5]}. As such, unlocking these insect collections is of paramount importance, and the large-scale nature of these collections necessitates the development of an efficient and effective digitisation process.

The 3D digitisation of specimens using current methods is either highly time-intensive or expensive, rendering it impossible to achieve across the collection in a reasonable time-frame. Yet, 3D models of specimens have huge potential for investigating species morphological responses to anthropogenic changes over time and identification of trade-offs in morphological responses within a 3D morphospace.

This project aims to develop a reproducible low-cost method of digitising specimens using commodified software to achieve large-scale efficient 3D digitisation of specimens. The student will experiment and develop the methods on the UMZC UK macromoth collection, and would gain experience in insect specimen handling and digitisation, as well as developing knowledge on the role of museum specimens in understanding the biodiversity crisis and tackling global challenges.

Some early experiments we have done with high quality mobile phones such as an iPhone 16 show that even off-the-shelf software (sucn as Polycam) using both the Lidar and just normal photogrammetry modes are sufficient to do a remarkably high fidelity 3D reconstruction of moths. The project, therefore, could either go in the direction of an app that uses ARKit to facilitate an interactive scan, or towards building a low-cost rig within which the insect mounting board could be placed with the camera going around "on rails". Challenges will include developing the photogrammetry software pipeline, and also on matters of focus to ensure that the critical areas are measured appropriately accurately (such as the antenna).

1 Project notes

Beatrice Spence, Arissa-Elena Rotunjanu and Anna Yiu ran this as a summer internship in 2025 under the name DoME (Digitisation of Museum Entomology), supervised by Tiffany Ki and Edgar Turner, and kept weeknotes throughout. The work fell into three strands: establishing what off-the-shelf capture software can actually do to a pinned insect, working out how to process the results at collection scale, and building a rig so that the capture is consistent and largely unattended.

The first investigation strand found that capture methods actually fail in quite complementary ways.

• Gaussian splatting gets the shape right (including transparent wings, legs and antennae) but overly smooths out the patterns that are important to preserve for insects.
• Photogrammetry captures the patterns and surface texture but thickens thin structures and tends to lose antennae and legs altogether.
• LiDAR handles fuzzy surfaces but is poor at anything thin (quite common in insects!).

For Gaussian splatting, more photographs do not reliably produce a better model, and a hundred well-chosen images sometimes beat a thousand taken in burst. Measurements taken from the resulting models also matched the group's existing ImageJ workflow closely without any calibration!

The processing was constrained by available hardware. A commercial Gaussian splatting API would cost on the order of £100k+ to run across the collection, so the team moved to training their own local models, working through COLMAP, nerfstudio and Brush.

We also started collaborating with Aurojit Panda and Hexu Zhao at NYU, whose Grendel-GS work on scaling up 3D Gaussian splatting training offers a viable route to conducting this work at a collection scale. Hexu helped us out with some very good reconstructions from the summer's data!

The rig itself is a white 3D-printed dome under half a metre across, printed in eight sections (because otherwise it wont fit on a 3D printer), with LED strips ringing each of four camera holes and a motorised turntable underneath driven by a stepper motor and a Raspberry Pi. The turntable rotates the dome and cameras around the specimen, rather than rotating the specimen itself. A 20th century pinned moth shouldn't be subjected to any more vibration than strictly necessary!

Arissa-Elena Rotunjanu wrote an iOS app that fires the shutter at fixed intervals so that the operator only has to position the phone, and Beatrice Spence wrote another that displays the captured models alongside their catalogue data.

Alex Ho, Michael Dales, Sadiq Jaffer and I gave design input on the dome and lighting along with Ed and many others from the museum. The dome was finished by the end of the project, and the collection cataloguing will hopefully continue!

1.1 Weeknotes

• Week 1, 30 June 2025. Specimen handling induction, and a first pass at Kiri Engine and Polycam on pinned moths. Ten attempts in a day established that a moving white plastazote background helps capture antennae, and that around a thousand images gives the best Gaussian splat for the time spent. First sketch of a dome with four camera positions.
• Week 2, 7 July 2025. Costed the commercial APIs, found them prohibitive, and settled on training a model instead. Varying the image count on a single Pine Hawk-moth showed that Gaussian splatting does not improve monotonically with more photographs. Transparent wings turned out to be the case only Gaussian splatting handles.
• Week 3, 14 July 2025. The scAnt scanner came online, giving a high-fidelity reference to validate the phone captures against. Reading on Gaussian splatting variants, and the first app prototypes in Swift and React Native.
• Week 4, 21 July 2025. Processed the scAnt captures into models on a desktop GPU after a laptop proved inadequate, added 3D viewing to the app, and opened the correspondence with NYU.
• Week 5, 25 August 2025. The dome design was too large for any accessible printer, so it was split into sections and prototyped at reduced scale. Work began on the interval-shooting camera app, and on getting a training pipeline onto a machine with a suitable GPU.
• Week 6, 1 September 2025. A design review led to a redesign: a smooth interior for a uniform background, lighting moved to sit around the cameras, and the diameter cut so that eight sections fit the printer bed. Also looked at whether Segment Anything has a 3D equivalent that could isolate body parts for measurement.
• Week 7, 8 September 2025. Printed a test section to check that the phone slots hold a range of handsets at the right angle, chose the LED strips on colour rendering index, and settled the turntable design on spinning the dome rather than the specimen.
• Week 8, 15 September 2025. Printed half the dome in time to present the design to the Insect Ecology and Agroecology groups, and assembled and soldered the LED lighting.
• Week 9, 22 September 2025. Laser cut the turntable plate, fixed the lighting permanently, and drove the stepper motor from a Raspberry Pi. The dome was completed, with the final panel held on by magnets so that specimens can be placed inside.
• Wrap-up, 26 September 2025. A summary of the whole project, and a list of what they would do next: parametrise the CAD so the dome can be resized independently of the phone slots, trigger the camera on stillness rather than on a timer, formally compare the three capture methods using data from the finished rig, and possibly take the rig into the field with a cooling system to scan live insects.