Hey there! I'm a researcher at the Computer Vision & Graphics group of Fraunhofer HHI in Berlin.
My interests are in understanding and reconstructing the world around us. I'm fascinated by strategies to compact higher dimensional spaces into lower ones, and competitive learning strategies.
Milena Bagdasarian, Paul Knoll, Yi-Hsin Li, Florian Barthel, Anna Hilsmann, Peter Eisert, Wieland Morgenstern2024-2025 · Eurographics 2025 STAR
We present a work-in-progress survey on 3D Gaussian Splatting compression methods, focusing on their statistical performance across various benchmarks. This survey aims to facilitate comparability by summarising key statistics of different compression approaches in a tabulated format. The datasets evaluated include Tanks and Temples, Mip-NeRF 360, Deep Blending, and Synthetic NeRF. For each method, we report the Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), Learned Perceptual Image Patch Similarity (LPIPS), and the resultant size in megabytes (MB, 10^6 bytes), as provided by the respective authors. This is an ongoing, open project, and we invite contributions from the research community as GitHub issues or pull requests.
Wieland Morgenstern, Florian Barthel, Anna Hilsmann, Peter Eisert2024 · European Conference on Computer Vision (ECCV), 2024
We propose a novel compact scene representation for 3D Gaussian Splatting by organizing parameters into a 2D grid and enforcing local smoothness during training. This approach leverages off-the-shelf image compression to efficiently store attribute images, reducing storage size by 19.9x to 39.5x while maintaining high visual quality. Our method includes an efficient GPU-based sorting algorithm and provides a simple interface for compressing and decompressing 3D scenes.
Wieland Morgenstern, Milena T. Bagdasarian, Anna Hilsmann, Peter Eisert2024 · IEEE Conference on Virtual Reality and 3D User Interfaces (IEEE VR), 2024
We propose a novel representation of virtual humans for highly realistic real-time animation and rendering in 3D applications. We learn pose-dependent appearance and geometry from accurate dynamic mesh sequences obtained from multiview-video reconstruction. By leveraging SMPL body models as a-priori knowledge, we learn the difference between observed geometry and the fitted SMPL model, encoding both pose-dependent appearance and geometry in the consistent UV space of the SMPL model. This approach ensures high realism while facilitating streamlined animation and rendering.
Wieland Morgenstern, Niklas Gard, Simon Baumann, Anna Hilsmann, Peter Eisert2023 · IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2023
We present a new approach to direct depth estimation for Spatial Augmented Reality applications using event cameras paired with laser projectors. Our key contribution involves converting projector time maps into rectified X-maps that capture x-axis correspondences for incoming events, enabling direct disparity lookup without additional search. This significantly simplifies depth estimation, making it more efficient while maintaining accuracy. Our method compensates for non-linear temporal behavior of laser projectors through time map calibration, resulting in improved performance. With depth estimation executed by just two lookups in under 3ms per frame, our approach enables real-time interactivity ideal for augmented reality experiences where low latency is crucial.
Wieland Morgenstern, Anna Hilsmann, Peter Eisert2019 · ACM SIGGRAPH European Conference on Visual Media Production (CVMP), 2019
We present an algorithm for progressive mesh registration to provide temporal consistency, simplify temporal texture editing and optimize the data rate of 3D mesh sequences recorded in a volumetric capture studio. Our approach splits sequences into groups of frames that share connectivity, using keyframes that are progressively deformed to approximate adjacent meshes. The algorithm employs a coarse-to-fine ICP approach that is robust against large deformations while preserving small details. A deformation graph constrains transformations to be locally as-rigid-as-possible, allowing work with any natural objects. Our method robustly tracks human actors with varying clothing over hundreds of frames, taking less than five seconds per frame on a single desktop machine.
Exploring the colors of moving pictures. I'm fascinated with colors in moving pictures, and started this project to visualize the colors and their change over time in films. It's quite hard to boil down all the information just one movie contains into a single image, but fun to play around with different and unique plots.
For my bachelor's thesis, I implemented the Histograms of Oriented Gradients algorithm highly parallel for GPUs, with OpenCL. To speed up the calculations, I developed a new algorithm for the block computation and as a result surpassed several other implementation.
An hour-long conversation with Michael Rubloff and MrNeRF in the View Dependent Podcast about 3DGS compression, why the Gaussians are Self-Organizing, and how this project came to be.
Eurographics 
