Jan Eric Lenssen Senior Researcher at Max Planck Institute for Informatics, Founding Engineer at Kumo.ai Short Biography Heading the Geometric Representation Learning Group at MPI-INF in D2: Computer Vision and Machine Learning. Founding team member of Kumo.ai (Stanford / TU Dortmund startup), bringing GNNs to relational data. ELLIS Unit Faculty Member and Saarland Informatics Campus Fellow. Dr.-Ing. in Computer Science (Feb. 2022) from TU Dortmund University under Heinrich Müller. Research internships at Facebook (Meta) Reality Labs and Nnaisense. [Google Scholar] [LinkedIn] [Github] [Email]

Research Interests

Selected Publications

• TokenFormer: Rethinking Transformer Scaling with Tokenized Model Parameters arXiv preprint Haiyang Wang, Yue Fan, Muhammad Ferjad Naeem, Yongqin Xian, Jan Eric Lenssen, Liwei Wang, Federico Tombari, Bernt Schiele Short Abstract: An alternative to the transformer architecture. Views trainable parameters as tokens to allow for incremental model scaling, leading to more efficient training. [Project Page] [Paper]

• Spurfies: Sparse Surface Reconstruction using Local Geometry Priors 3DV 2025 (accepted) Kevin Raj, Christopher Wewer, Raza Yunus, Eddy Ilg, Jan Eric Lenssen Short Abstract: A method leveraging synthetic data to learn local surface priors for surface reconstruction from few images. Can be applied to both bounded and unbounded scenes. [Project Page] [Paper]

• InterTrack: Tracking Human Object Interaction without Object Templates 3DV 2025 (accepted) Xianghui Xie, Jan Eric Lenssen, Gerard Pons-Moll Short Abstract: A tracker for dynamic humans and objects under occlusion from a monocular RGB video without object templates. Trained on synthetic data and generalizing to in-the-wild videos. [Project Page] [Paper]

• Scribbles for All: Benchmarking Scribble Supervised Segmentation Across Datasets NeurIPS 2024 - Datasets and Benchmarks Wolfgang Boettcher, Lukas Hoyer, Ozan Unal, Jan Eric Lenssen, Bernt Schiele Short Abstract: A set of scribble-labeled semantic segmentation datasets and an algorithm to automatically obtain such labels for a densely labeled dataset. [Project Page] [Paper]

• RelBench: A Benchmark for Deep Learning on Relational Databases NeurIPS 2024 - Datasets and Benchmarks Joshua Robinson, Rishabh Ranjan, Weihua Hu, Kexin Huang, Jiaqi Han, Alejandro Dobles, Matthias Fey, Jan Eric Lenssen, Yiwen Yuan, Zecheng Zhang, Xinwei He, Jure Leskovec Short Abstract: An open benchmark for machine learning on relational databases. Contains a collection of realistic, large-scale, and diverse benchmark datasets, a leaderboard and unified evaluation. [Project Page] [Paper]

• From Similarity to Superiority: Channel Clustering for Time Series Forecasting NeurIPS 2024 Jialin Chen, Jan Eric Lenssen, Aosong Feng, Weihua Hu, Matthias Fey, Leandros Tassiulas, Jure Leskovec, Rex Ying Short Abstract: A channel clustering strategy for multivariate time series forecasting that allows to weight different expert models for different time series channels. [Paper]

• Improving 2D Feature Representations by 3D-Aware Fine-Tuning ECCV 2024 Yuanwen Yue, Anurag Das, Francis Engelmann, Siyu Tang, Jan Eric Lenssen Short Abstract: We show that finetuning 2D foundation models with descriptors that have been fused into a 3D Gaussian representation improves feature quality for downstream tasks. [Project Page] [Paper]

• latentSplat: Autoencoding Variational Gaussians for Fast Generalizable 3D Reconstruction ECCV 2024 Christopher Wewer, Kevin Raj, Eddy Ilg, Bernt Schiele, Jan Eric Lenssen Short Abstract: A fast autoencoder that encodes image pairs into 3D variational feature Gaussians that model uncertainty individually for different locations in 3D space. Trained purely on videos. [Project Page] [Paper]

• Neural Parametric Gaussians for Monocular Non-Rigid Object Reconstruction CVPR 2024 Devikalyan Das, Christopher Wewer, Raza Yunus, Eddy Ilg, Jan Eric Lenssen Short Abstract: A two-stage approach for non-rigid reconstruction from monocular videos. A coarse point template is found first to act as regularization for a local 3D Gaussians representation. [Project Page] [Paper]

• NRDF: Neural Riemannian Distance Fields for Learning Articulated Pose Priors CVPR 2024 (Highlight) Yannan He, Garvita Tiwari, Tolga Birdal, Jan Eric Lenssen, Gerard Pons-Moll Short Abstract: A method to learn data priors for articulated poses (humans, hands, animals) using a Riemannian distance field formulation and Riemmanian gradient descent. [Project Page] [Paper]

• Template Free Reconstruction of Human-object Interaction with Procedural Interaction Generation CVPR 2024 (Highlight) Xianghui Xie, Bharat Bhatnagar, Jan Eric Lenssen, Gerard Pons-Moll Short Abstract: A hierarchical diffusion model for human-object interaction reconstruction from single image, trained on ProciGen, a large-scale, automatically created, synthetic dataset. [Project Page] [Paper]

• Neural Point Cloud Diffusion for Disentangled 3D Shape and Appearance Generation CVPR 2024 Philipp Schröppel, Christopher Wewer, Jan Eric Lenssen, Eddy Ilg, Thomas Brox Short Abstract: A denoising diffusion method on point clouds with features defining a radiance field. The disentangled representation allows for individual generation of shape and appearance. [Project Page] [Paper]

• GEARS: Local Geometry-aware Hand-object Interaction Synthesis CVPR 2024 Keyang Zhou, Bharat Bhatnagar, Jan Eric Lenssen, Gerard Pons-Moll Short Abstract: A method to generate realistic grasping sequences of hands, given wrist trajectory and object. We use local sensors on hand joints to obtain features independent of global object information. [Project Page] [Paper]

• Relational Deep Learning: Graph Representation Learning on Relational Databases ICML 2024 Position Papers Matthias Fey, Weihua Hu, Kexin Huang, Jan Eric Lenssen, Rishabh Ranjan, Joshua Robinson, Rex Ying, Jiaxuan You, Jure Leskovec Short Abstract: A position paper introducing the concept of relational deep learning, an end-to-end framework to learn directly on relational databases using graph neural networks. [Project Page] [Paper]

• Recent Trends in 3D Reconstruction of General Non-Rigid Scenes Eurographis State of the Art Reports 2024 Raza Yunus, Jan Eric Lenssen, Michael Niemeyer, Yiyi Liao, Christian Rupprecht, Christian Theobalt, Gerard Pons-Moll, Jia-Bin Huang, Vladislav Golyanik, Eddy Ilg Short Abstract: A state of the art report for 3D reconstruction of general non-rigid scenes, covering 3D and 4D representations, deformation models, generalizable reconstruction, automatic decomposition and editing. [Project Page] [Paper]

• SimNP: Learning Self-Similarity Priors between Neural Points ICCV 2023 Christopher Wewer, Eddy Ilg, Bernt Schiele, Jan Eric Lenssen Short Abstract: A renderable neural point radiance field that learns category-level self-similarities from data by connecting coherent neural points to embeddings via optimized bipartite attention scores. [Project Page] [Paper]

• Pose-NDF: Modeling Human Pose Manifolds with Neural Distance Fields ECCV 2022 (Oral Presentation, Best Paper Honorable Mention Award) Garvita Tiwari, Dimitrije Antic, Jan Eric Lenssen, Nikolaos Sarafianos, Tony Tung, Gerard Pons-Moll Short Abstract: An unsigned neural distance field that models the manifold of plausible human poses in high-dimensional SO(3). Given human poses can be projected onto the manifold by SO(3) gradient descent. [Project Page] [Paper] [Code]

• TOCH: Spatio-Temporal Object-to-Hand Correspondence for Motion Refinement ECCV 2022 Keyang Zhou, Bharat Lal Bathnagar, Jan Eric Lenssen, Gerard Pons-Moll Short Abstract: A spatio-temporal representation of point-wise correspondences between a parameterized hand template mesh and an object mesh, which allows for refining captured grasping motions. [Project Page] [Paper] [Code]

• GnnAutoScale: Scalable and Expressive Graph Neural Networks via Historical Embeddings ICML 2021 Matthias Fey, Jan Eric Lenssen, Frank Weichert, Jure Leskovec Short Abstract: A framework to scale arbitrary message passing graph neural networks to large input graphs using historical embeddings. The scaling method provably preserves the GNN expressiveness. [Paper] [Code]

• Quaternion Equivariant Capsule Networks for 3D Point Clouds ECCV 2020 (Oral Presentation) Yongheng Zhao, Tolga Birdal, Jan Eric Lenssen, Emanuele Menegatti, Leonidas Guibas, Federico Tombari Short Abstract: A provably SO(3)-equivariant capsule network architecture to classify and canonicalize point clouds. Introduces a routing mechanism based on the iterative least-squares Weiszfeld algorithm. [Paper] [Code]

• Deep Local Shapes: Learning Local SDF Priors for Detailed 3D Reconstruction ECCV 2020 Rohan Chabra, Jan E. Lenssen, Eddy Ilg, Tanner Schmidt, Julian Straub, Steven Lovegrove, Richard Newcombe Short Abstract: A voxel grid of implicit neural fields as local priors for representing signed distance functions. Enables detailed 3D reconstructions of full scenes while only using synthetic training data. [Paper]

• Deep Iterative Surface Normal Estimation CVPR 2020 (Oral Presentation, Best Paper Nominee) Jan Eric Lenssen, Christian Osendorfer, Jonathan Masci Short Abstract: A differentiable, iterative re-weighted least-squares (IRLS) algorithm for normal estimation on unstructured point clouds. The method combines the efficiency of traditional IRLS with a GNN data prior. [Paper] [Code]

• Deep Graph Matching Consensus ICLR 2020 Matthias Fey, Jan Eric Lenssen, Christopher Morris, Jonathan Masci, Nils Kriege Short Abstract: A two-stage siamese graph neural network for graph matching in several applications. Refines local feature matchings by optimizing neighborhood consensus in the second stage. [Paper] [Code]

• Weisfeiler and leman go neural: Higher-order graph neural networks AAAI 2019 Christopher Morris, Martin Ritzert, Matthias Fey, William Hamilton, Jan E. Lenssen, Gaurav Rattan, Martin Grohe Short Abstract: Shows equivalence in expressiveness of specific graph neural networks and the Weisfeiler-Leman graph isomorphy test. Introduces higher order GNNs with provably more expressiveness. [Paper] [Code]

• Fast Graph Representation Learning with PyTorch Geometric ICLR 2019 Workshop Matthias Fey, Jan Eric Lenssen Short Abstract: Heavily used PyTorch-based library to write and train graph neural networks for several applications. Implements efficient, differentiable and customizable message passing on GPU and CPU. [Website] [Paper] [Code]

• Group Equivariant Capsule Networks NeurIPS 2018 Jan Eric Lenssen, Matthias Fey, Pascal Libuschewski Short Abstract: A provably group equivariant capsule network for rotation invariant classification and pose estimation. Introduces an equivariant dynamic routing algorithm for images. [Paper] [Code]

• SplineCNN: Fast geometric deep learning with continuous B-spline kernels CVPR 2018 Matthias Fey*, Jan Eric Lenssen*, Frank Weichert, Heinrich Müller Short Abstract: A differentiable operator for continuous convolution on irregular-structured data. Parameterizes a continuous kernel using B-splines and provides efficient GPU implementations via message passing. [Paper] [Code]

• Real-time Low SNR Signal Processing for Nanoparticle Analysis with Deep Neural Networks BIOSIGNALS 2018 (Best Paper Award) Jan Eric Lenssen, Anas Toma, Albert Seebold, Victoria Shpacovitch, Pascal Libuschewski, Frank Weichert, Jian-Jia Chen, Roland Hergenröder Short Abstract: A multistage CNN architecture for detecting low-SNR particles in images from a surface plasmon resonance sensor. The method performs detection, classification and size estimation of nanoparticles. [Paper]