Visualising & Modelling Trees
Tasmania has some of the largest trees in the world – giants that can be thousands of years old, growing in its not-so-remote temperate rainforests. This is an experiment in making a reasonably accurate photogrammetric model of a very large Eucalyptus Regnans – sadly killed by the devastating bushfires around the Huon Valley and Geeveston in 2019.
Trees are tricky for 3D modelling – especially if shooting from ground level in a busy forest environment. As a first approach I’ve shot the tree on a cloudy day using a 50mm lens, attempting to regularise my distance from the tree – with varying success as one clambers over logs, sinks into the ground or is simply stopped by impenetrable growth. Peering up towards the tree confuses the reconstruction algorithm, mixing points that should be ‘infinitely’ distant (the sky) and those of upper tree branches – essentially a problem of parallax. However, we can clearly discern branch structure within this indistinct cloud. After processing the photos by assessing quality (e.g. focus) and optimising for view position and overlap, the pointcloud cleans up a bit and can be further cleaned with masking and quite a bit of manual work. At this point the main trunk structure and lowest branches are fairly well distinguished and can be used to create high-quality polygon meshes. I’ll add some renders sometime.
Next steps to consider will be training a classifier in CloudCompare (CANUPO) to separate the branches and sky – this looks promising. I also wonder what sort of new machine learning approaches would be of use – one can imagine a neural net approach that might be trained to distinguish branching structures and also to complete the upper surface of branches that cannot be photographed from ground level, using either learned parameters and hallucinating the top surfaces or estimating some kind of surface of revolution. Of course, this sort of inferred completion would be mostly useful for creating manifolds for more informal applications such as game environments, 3D scenes and printed models. A good start.