Analysis and Prediction of Deforming 3D Shapes using Oriented Bounding Boxes and LSTM Autoencoders

Deforming 3D shapes that can be subdivided into components can be found in many different areas, for example crash analysis, structure, and material testing, or moving humans and animals that can be subdivided into body parts. The data can be especially challenging since it has a temporal and a spatial dimension that have to be considered jointly. In particular, we consider shape data from Computer Aided Engineering (CAE), where numerical simulations play a vital role in the development of products, as they enable simpler, faster, and more cost-effective investigations of systems. If the 3D shape is complex, the detection of patterns in the deformation, called deformation modes in CAE (figure 1), speeds up the analysis. While deformation modes are dependent on model parameters, their behavior is accessible, albeit tediously, upon manifestation when having access to the completed simulation run. In practice this detection of deformation modes is based on error-prone supervision of a nontrivial, hand-selected subset of nodes in critical components. In general, simulation results contain abundant features and are therefore high dimensional and unwieldy, which makes comparative analysis difficult. Intuitively, the sampled points describing the component are highly correlated, not only in space but also over time. This redundancy in the data invites further analysis via feature learning, reducing unimportant information in space as well as time and thus dimensionality.