Deep learning on 3D data is still challenging — partly due to the computational complexity of training on 3D grids. Thus, alternative data structures become more and more interesting. PointNets, for example, operate directly on unordered set of points, i.e. point clouds. Similarly, Point Set Generation Networks are able to directly predict point clouds from images. In this article, I briefly summarize both ideas.
This article presents visualizations of the qualitative and quantitative results from the superpixel benchmark published in CVIU. Based on NVD3 and Unite Gallery, the visualizations allow to interactively compare different superpixel algorithms in the browser.
In April, I visited Prof. Bernt Schiele’s Computer Vision and Multimodal Computing Department at the Max Planck Institute for Informatics in Saarbrücken. Aside from a presentation on my recent superpixel benchmark, I also met many interesting people and learned a lot about a career in research.
Lifting convolutional neural networks to 3D data is challenging due to different data modalities (videos, image volumes, CAD models, LiDAR data etc.) as well as computational limitations (regarding runtime and memory). In this article, I want to summarize several recent papers addressing these problems and tackling different applications such as shape recognition, shape retrieval, medical image segmentation or object detection.