In this CVPR’18 paper, based on my master thesis, we propose a weakly-supervised and learning-based approach to 3D shape completion of sparse and noisy point clouds. We show that, using a learned shape prior, shape completion can be learned without access to ground truth shapes — only by knowing the object category at hand. This article provides the paper and its supplementary material.
My master thesis, written at the Autonomous Vision Group of Max Planck Institute for Intelligent Systems under the supervision of Prof. Andreas Geiger, addresses the problem of 3D shape completion of sparse point clouds under weak supervision. Specifically, based on a learned shape prior it is possible to learn 3D shape completion without access to ground truth shapes, as shown on KITTI. This article briefly introduces the problem and the main contributions and offers the thesis as download.
As part of the online course Creative Applications of Deep Learning with TensorFlow, and to get started with TensorFlow, I implemented some experiments on MNIST. Specifically, I tested different architectures, activation functions and initialization schemes. While these experiments are not systematic enough for reliable results, they can be useful as an introduction to TensorFlow. In this article, I want to share the code and the corresponding presentation.
In the last few months, I started to pursue a PhD. Although I did not have many options, also because I decided not to apply to many programs, I found choosing the right PhD incredible difficult. In this article, I want to share some of my insights.
Recently, I started reviewing for different conferences and journals. Based on my previous work, specifically the Superpixel Benchmark, I mostly reviewed papers on superpixel algorithms. For most of these algorithms, and additional related work, I made notes; in this article I want to keep track of the algorithms and benchmarks. The goal is to have an up-to-date list of superpixel algorithms and their implementations.
In computer graphics, watertight meshes usually describe meshes consisting of one closed surface. In this sense, watertight meshes do not contain holes and have a clearly defined inside. Therefore, they are commonly required by many applications in computer graphics as well as in computer vision — for example, when voxelizing meshes into occupancy grids or signed distance functions. However, I found it very difficult to find a proper formal definition of watertightness. In this article, I want to discuss the definition I used for my master thesis.
In the last few days I took some time to revisit some of my CMSimple plugins: News, a news article and blog plugin, Pictures, an image gallery plugin, YouTube, a plugin to create YouTube galleries, and BBClone, a plugin for web analytics with BBClone. In this article, I want to present the main features, the updated documentation as well as two demo applications to try out the plugins.
Part of my master thesis at the Max Planck Institute for Intelligent Systems was an initial proposal — outlining the general idea and the current state-of-the-art. Specifically, I worked on learning 3D shape completion on KITTI using 3D bounding boxes only. In this article, I want to present this proposal.
As part of my stay in the Autonomous Vision Group in Tübingen, I also worked on the KITTI dataset. With the help of Bo Li, we were able to add a novel benchmark: the 3D object detection benchmark. In this article, I briefly want to introduce the benchmark and give some useful hints regarding submission.
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.