Thesis title: Deep Learning on 3D Point Clouds: Methods and Perspectives
Shape correspondence and completion are cornerstone problems in the field of computer vision and have historically attracted much research activity.
Recently, there has been a noticeable change in the environmental condition in which it arises. The quantity of data available has increased drastically, thanks to the ubiquity of acquisition devices, even at the user grade level. New field of applications have grown in importance, such as virtual reality and self driving cars, mostly due to the increased in computational capacity brought by the new and improved hardware. These changing conditions call for new method able to keep up with this new state. Deep learning seems to answer this call, with techniques developed in surrounding context, spilling over to computer vision.
It is in this context that we propose novel data driven deep learning techniques able to effectively tackle the problem of shape correspondence. Specifically, we first propose an alignment-based solution to the problem of shape completion from range scans. Our data-driven solution is based on learning the space of distortions, linking scans at various poses to whole shapes in other poses. As a result, at test time we can accurately align unseen pairs of parts and whole shapes at different poses.
Additionally, we propose the first transformer based architecture to tackle the problem of non–rigid registration. We introduce a novel surface attention mechanism better suited to exploit the local geometric priors of the underlying structure. Our method reaches state of the art performance in shape matching and shape registration without assuming any fixed template, and generalizes also to different and complex geometries.