Projects Last Projects

We propose a metric learning framework for the construction of invariant signatures of non-rigid 3D point clouds under the isometry transformations group. We leverage the representational power of convolutional neural networks to compute these signatures and show that in comparison with classical methods, we achieve superior results that allow for higher classification accuracy using the invariant signature, and a lower pose dependency, with the additional advantage of much lower complexity, allowing for the calculation of invariant signatures for larger point clouds with orders of magnitude less time, this is achieved without the use of edge information that is commonly used for such applications. Furthermore, our proposed training scheme allows achieves superior classification accuracy than end-to-end trained classifiers using the same architecture.

Please, see project report.

Please, see final presentation.