Embedded Files
My research focuses on the interplay of geometric modeling, shape analysis, and machine learning, and how they can be used to create intelligent systems. With a focus on Deep Shape Representation for Shape Analysis, Modeling, and Reconstruction, my goal is to utilize unsupervised or self-supervised learning methods to automatically analyse the composition and structure of geometric objects or generate new ones.
Currently, I research neural networks for geometry abstraction. Furthermore, I supervise multiple undergraduate students developing methods for 3D shape representation, generation, partial symmetry detection as well as 3D structure and relationship analysis. My teaching obligations include the lectures on "Shape Analysis and 3D Deep Learning" as well as "Data Analysis and Visualization".
Keywords: Geometric Modeling, Shape Analysis, Machine Learning, Generative Models, 3D Deep Learning, Computer Vision, Computer Graphics
2024
Learning Fine-to-Coarse Cuboid Shape Abstraction
G. Kobsik, L. Kobbelt, ongoing research
Abstracting geometric shapes simplifies complex objects, enabling efficient downstream tasks such as co-segmentation, structural analysis or symmetry detection. We solve this problem using Transformers and cuboids. First, we reconstruct the 3D shape using an extensive amount of primitives, and second we restrict their number to represent only the most defining parts of the object.
2023
Partial Symmetry Detection for 3D Geometry using Contrastive Learning with Geodesic Point Cloud Patches
G. Kobsik, I. Lim, L. Kobbelt, 2023, arXiv pre-print
Partial or local symmetry detection remains an ambigious problem in computer graphics and shape analysis. A research domain without any groundbreaking contributions since 2006, nor AI-based approaches. We addressed this challange by learning to compare local geometric patches and identifying clusters of similar patches in the latent space as potential symmetric regions.
Octree Transformer: Autoregressive 3D Shape Generation on Hierarchically Structured Sequences
M. Ibing, G. Kobsik, L. Kobbelt, 2023, Conference on Computer Vision and Pattern Recognition (CVPR) - StruCo3D Workshop
Based on the findings of my master thesis, we formulated a fully autoregressive decoding scheme for the transformer architecture using octrees to generate 3D shapes. We showed that this approach is a viable alternative to current generative models and able to generate a wide variety of diverse shapes, both conditionally and unconditionally on the ShapeNet dataset.
2021
Shape Generation utilizing Octrees with Self-Attention Networks
G. Kobsik, 2021, Master Thesis, RWTH Aachen University
I researched different formulations of autoregressive self-attention encoder and decoder neural networks with the goal of generating 3D shapes. Especially, I formulated the shape generation task as a sequence generation task on a linearized octree structure and thus was able to apply SOTA natural language processing models to the field of shape generation, e.g. the transformer architecture. A remarkable property of the formulation is the ability to inherently define voxel super-resolution as a sequence completion task.
Development of a Heuristic-based Agent for an Interactive Multiplayer Game
G. Kobsik, R. Kupper, M. Pozor, 2021, Competition Submission, informatiCup 2021
This project is the contribution to the informatiCup challange 2021 of the German Informatics Society (GI). We developed our own heuristic-based AI agent that is able to competitively play the game of spe_ed. Our submission was selected among many others for the final round of the contest, in which we placed 1st.
2019
Aided Hand Detection in Thermal Imaging using RGB Stereo Vision
M. Schmieschek, G. Kobsik, A. Stollenwerk, S. Kowalewski, T. Orlikowsky, M. Schoberer, 2019, 41st Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC)
During my bachelor studies, I was able to support the development of a contact-free hand desinfection validation application based on a thermal camera combined with two RGB cameras. Our goal was to develop a system, which is able to detect and segment a hand as robustly as possible to detect temperature gradients before and after the desinfection of hands and thus support the medical personal in their daily desinfection routine to prevent the spread of infections in clinical environments.
2018
Development of a Real-time Dynamic Rigidbody Fracture Engine in Unity
G. Kobsik, B. Beqa, 2018, Practicel Course
As part of a practical course of the Visual Computing Institute we implemented a simple fracture engine for Unity based on the paper "Real Time Dynamic Fracture with Volumetric Approximate Convex Decompositions" by Müller et al. The developed engine allows to add dynamic fracture behaviour to rigid bodies by simply adding a single script to an object. No add-ons nor other frameworks need to be installed within Unity as everything is handled by the implemented code.
Foam Generation for Particle-based Fluid Simulations
G. Kobsik, 2018, Bachelor Thesis, RWTH Aachen University
My first scientific research focused on the generation and advection of foam particles in fluid environments to enhance the visual appearance of the simulations. Based on observations and a simple approximation of the underlying physical behaviour, I was able to depict different interactions of air with water to add physically-based effects like foam, bubbles or spray in dynamic scenes.
Development of a Lego Mindstorms NXT Segway Roboter based on a PID-controller
J. Breyer, F. Friedrichs, C. Kloos, G. Kobsik, R. Kupper, 2018, Practical Course
Together in a small team we developed a self-balanced roboter based on the Lego Mindstorms toolkit. Given only very basic framework and imprecise hardware sensors, we managed to built a roboter which is balanced by a virtual PID-controller, two wheels and an acceleration sensor. With further add-on, like game pad steering, virtual environment mapping and coordinate based controlling, we solved different tasks in navigating a small parcour.
2017
Development of a Dynamic Grassland Generator in Unity
G. Kobsik, T. Jochum, 2017, Game Development Course
One of my first bigger practical projects included the development of a grassland simulation in Unity. We managed to implement different styles of grass represented by billboards or procedurally generated geometry on the GPU. Furthermore, we implemented dynamic environmental effects like rain and wind. The skydome is a physically-based simulation of the interaction between the atmosphe and the sunlight, approximating effects like Rayleigh and Mie scattering. Our goal was to learn to utilize different shaders (geometry, tesselation, compute, etc.) provided by the GPU pipeline.
READ REPORT
Page updated
Google Sites
Report abuse