Welcome to the website of the Geometric Representation Learning (GRL) group within department D2: Computer Vision and Machine Learning of the Max-Planck-Institute for Informatics.
Our research focus is on representations and algorithms for perception and inference in the 3D world, putting us at the intersection of computer vision, computer graphics and machine learning.
Dr. Jan Eric Lenssen
Senior Researcher
Research Vision
Computer vision becomes increasingly capable of representing the 3D world, given large sets of complete observations from sensors such as cameras or 3D scanners. However, in comparison with humans, we still lack an important ability: deriving complete representations from incomplete observations. Our group works towards replicating this human ability to approach the desired capabilities of a general computer vision system.
A key aspect of the human abilities is that observations we make are complemented by previously learned information: the world is not only sensed - to a large degree it is inferred. As a consequence, we approach this task by developing efficient machine learning algorithms that compress information from large datasets and use that information to perform inference.
We tackle algorithmic research topics in the areas of:
- Efficient neural fields for volumetric representation
- Processing of irregular structured and sparse data
- Generative models for spatial reasoning
To advance the fields of:
- 3D reconstruction from incomplete observation
- Object- and scene generation
- Scene understanding and reasoning
If you are interested to be part of the team, reach out to us directly or via the D2 application options! We offer PhD positions. We also have Master thesis topics and HiWi positions for students from University of Saarland. We do not offer short-term internships for people not already at University of Saarland.
News
Saarland/Intel Program
February 2024
Our project "Tractable Diffusion Models for Large-Scale 3D Scene- and Object-level Representations" was accepted and funded by the the Saarland/Intel Joint Program on the Future of Graphics and Media. We are looking forward to a successfull collaboration, in which we aim to bring generative models to efficient 3D representations!
SimNP accepted at ICCV 2023
July 2023
Happy to announce that our paper SimNP: Learning Self-Similarity Priors between Neural Points was accepted to ICCV 2023. We present a category-level, generalizable neural point representation that learns similarity between neural points as a prior for single- and few-view reconstruction.
Best Paper Honourable Mention at ECCV2022!!
Oct 2022
Congratulations to Garvita Tiwari, Dimitrije Antic, Jan Eric Lenssen, Nikolaos Sarafianos, Tony Tung and Gerard Pons-Moll for receiving the Best Paper Honorable Mention at ECCV'22 for the paper Pose-NDF: Modeling Human Pose Manifolds with Neural Distance Fields. 3 paper awards were given out of 6773 submissions.
Latest Publications
Spurfies: Sparse-view Surface Reconstruction using Local Geometry Priors
in International Conference on 3D Vision (3DV), 2025.
InterTrack: Tracking Human Object Interaction without Object Templates
in International Conference on 3D Vision (3DV), 2025.
TokenFormer: Rethinking Transformer Scaling with Tokenized Model Parameters
in arXiv preprint, 2024.