In my Master’s Thesis, I focussed on point cloud processing architectures. I developed a set of intuitive benchmarks of robustness to rotations, and evaluated different data augmentation strategies, and a particular rotation-robust deep architecture based on a spherical lattice. I also proposed and evaluated an extension of that architecture, that helps collect information from the entire point cloud, rather than just near the spherical boundary.
You can download my Thesis here.
I worked on my Thesis in Spring/Summer 2020, at Prof. Maurice Fallon’s Dynamic Robot Systems group at the Oxford Robotics Institute, and had a very good time there!
Abstract:
In recent years, several different approaches have been developed, to obtain deep point cloud processing models that are robust or invariant to 3D rotations. This area of study is still evolving very rapidly and there is no consensus on the best approach. In this work, we focus on two independent subjects in the area.
Firstly, we compare different methods of using data augmentation to encourage rotation robustness in training. Related research works usually either use a single method of augmentation, without many remarks, or set a goal of avoiding data augmentation altogether. We design three easy-to-understand benchmarks of rotation robustness and use them to quantitatively compare multiple augmentation techniques. Interesting empirical observations about the benchmark scores are presented as conclusion.
Secondly, we investigate an existing rotation-robust architecture, the Spherical Fractal Convolutional Neural Network, in detail. We analyse its theoretical and empirical properties and propose a number of modifications to the model, that are evaluated on a standard object classification dataset. We demonstrate that our modifications yield a small improvement on established classification benchmarks.