05 Aug, 2021
Recently, our R&D team was tasked with the objective to take our 3D modelling capabilities to the next level, which led us to look at greater ways to solve real-world problems through our world-leading 3D data.
Aerometrex has established itself as an international leader in 3D modelling with clients such as Google, Microsoft, and many Australian government departments across local, state, and federal levels. With the help of our fleet of aircraft and hired helicopters, we capture aerial images of built areas that are translated into 3D models of cityscapes with details down to two centimetres per pixel. These models are a key tool for urban design and critical infrastructure projects, forming the basis for digital twins and smart city plans. As an example, government agencies use our 3D data as a crucial base model and add various other development assets, such as buildings, bridges, roads, etc. to assess and view them in context of their surroundings. This not only helps planning and design, but also gives them the visual tools they need for public consultation. 3D city models are being used extensively for analyses such as shadow casting, lines of sight and other similar features that are crucial to new developments within built environments. The tool’s use does not stop there; it can also be applied to future stages of construction and engineering plans.
Revolutionising 3D with Artificial Intelligence and Deep Learning
Our 3D models contain rich data & visual information. In the past, our R&D team has effectively identified various objects within these 3D models using semi-automatic processes, however extracting individual objects and deriving actionable information is time-consuming and not always scalable. The aim, thus, was to build a scalable solution that could use the underlying imagery to identify different objects within the 3D model. Individually labelling each pixel, would have taken extensive time & money. To solve the issue, the team developed a proprietary workflow to automatically produce labelled training datasets far more efficiently.
Our Geospatial Innovation Manager, Fabrice Marre says, “Deep learning was identified as the ideal tool for automating and implementing the process. We began to evaluate partners to work with us for further 3D development - someone with the right people & equipment to do this efficiently. Artificial intelligence (AI) would help save us time and deep learning would drive efficiency and far greater results. This automation would also be able to identify more objects and add critical attributes to them – for example, number of floors in a building, building volume, roof material, height of surrounding trees, etc."
Based on the defined requirements, the company identified Australian Institute for Machine Learning (AIML), as the ideal partner to support us in our endeavours. AIML ranks third in the world for machine learning expertise - with the expertise, knowledge, and computing power that could help supplement our in-house expertise and fast-track the transformation of our 3D models. Through these joint efforts, we are confidently working to build 3D models composed not just of shapes in RGB but of information – the detailed data within those shapes.
"Because Aerometrex already had so much information, ‘teaching’ the algorithm was relatively straightforward,” says Sam Hodge, Machine Learning Engineer - the AIML lead working with Aerometrex on the project. Once the teams have ‘taught’ the algorithm enough, we no longer need to label every object; the algorithm will keep learning for itself, drawing context and information not only from individual pixels but those surrounding it, and the tens of thousands of other pictures it has examined.", he adds.
Aerometrex has plans to keep growing and expanding the company’s services. The company continues to produce extensive data and transform it into actionable information. We have remained at the cutting edge of global 3D modelling and through our artificial intelligence and deep learning project, we are taking our solution to the next level. This AIML-partnered phase is the beginning of a long-term collaboration, one built on the confidence that, by partnering with the best in the industry, we can keep pushing the boundary. With AI and machine learning, we can offer users of 3D models more information about what’s within that model.