Generating 3D point clouds of Māori stone artefacts for machine learning training

Dr Mike laverick, Senior Solutions Specialist, Centre for eResearch

Reconstruct broken cultural artefacts

The project was a collaboration between Archaeology, Bioengineering and CeR to address a critical need for heritage consultants, iwi groups, museum curators and government heritage to generate high quality information retrieved from heritage mitigation work. The aim was to provide a technological solutions to significantly reduce the cost of heritage compliance and enable wider participation in archaeology by building 3D modelling, and shape analysis patents together with experience in the application of machine learning.

The role of the Centre for eResearch

CeR helped to develop high-throughput automated workflows from extremely high-fidelity 3D scans which were generated from thousands of Māori stone artefacts, culminating in 10TBs of raw data, with the goal of developing machine learning classification algorithms.

The automated workflow

We leveraged most major research platforms at the University to help process the data in time for the Ministry of Business, Innovation & Employment (MBIE) project deadline. The University’s Research Drive was used for data storage; the Nectar Research Cloud was deployed for compute power to leverage three A100 GPUs; and Globus was used to coordinate both internal University of Auckland data transfers and to automatically pull data from collaborators at the University of Otago (see Diagram 2).

In addition to automated workflows the Centre also helped develop processing pipelines to take raw scan data and convert it into machine-learning ready formats. The processing scripts clean the raw data of noisy data points and instrument artefacts and then perform artefact surface segmentation ready for training.