ANNs have become an established tool for data-driven approaches to all kinds of classification, regression and generative tasks. While impressive results have been achieved for image, video and other data modalities with a canonical vectorization, the performance for geometric shapes is still limited. This performance gap has a number of reasons including the difficulty to vectorize geometric shapes and the relatively small size of training datasets.
We have systematically explored various shape representations and evaluated their performance for a number of different ANNs structures in several classification as well as generative tasks. In my talk I will give an overview on the range of geometric data structures that we have investigated and present some results and insights that have emerged from these experiments. This includes the definition of local shape descriptors, the synthesis of procedural models as well as an effective auto-encoder for point clouds.
Back to Workshop II: Shape Analysis