Manuel Kaufmann

I am a computer scientist with a background in human-centric 3D Computer Vision. I previously worked as a tech lead in the executive team of the ETH AI Center and as postdoc at the AIT Lab at ETH Zurich, Department of Computer Science. I obtained my PhD in the AIT Lab under superivsion of Prof. Otmar Hilliges, co-advised by Prof. Markus Gross. I have completed my MSc at ETH Zurich with Prof. Hilliges as well and earned my BSc in Computer Science from the University of Basel under the supervision of Prof. Thomas Vetter.

I taught Machine Perception at ETH Zurich. All lecture materials including video recordings from 2025 are publicly available.

Scholar  /  LinkedIn

Research

I am interested in human-centric multimodal data acquisition pipelines, 3D human pose estimation from wearable sensors (e.g., IMUs and electromagnetic sensors) and single or multi-view images, 3D reconstruction of human avatars from images and video, and human motion modelling. For a selection of publications, see below.

A compositional 3D Gaussian avatar representation that enables free-form virtual try-on of clothing from multiple sources.

Personalized 3D human body fitting using point diffusion for improved accuracy in body pose and shape reconstruction.

An online Gaussian Splatting-based system for dense 3D reconstruction of both humans and their surrounding scenes from monocular video input.

An egocentric-inertial system that jointly performs human motion capture, localization, and dense scene mapping online.

Reconstructing humans wearing loose-fitting garments from in-the-wild monocular videos, handling challenging clothing dynamics.

A large-scale dataset captured from 2022 FIFA World Cup broadcasts for benchmarking global 3D human pose estimation in sports scenarios.

A holistic approach to jointly reconstructing humans and their surrounding 3D scene from a single monocular video.

Reconstructing multiple interacting people from a single monocular in-the-wild video, handling occlusions and close contact.

A new dataset of global 3D human pose and shape in-the-wild, captured using electromagnetic sensors for accurate ground truth.

A large-scale dataset capturing dexterous bimanual manipulation of articulated objects, useful for hand-object interaction research.

Expressive human avatars that capture body pose, facial expressions, and hand gestures in a unified model.

A dataset and method for 4D instance segmentation of two people engaged in close physical interaction.

A spatio-temporal transformer architecture that jointly models spatial and temporal dependencies for 3D human motion prediction.

3D human pose estimation from sparse wireless electromagnetic trackers.

Convolutional autoencoders for filling in missing frames in human motion sequences, enabling cleanup of noisy or incomplete motion capture data.

Incorporating the skeletal structure of the human body into the output layer improves 3D human motion modelling.

Real-time 3D human pose reconstruction from only six sparse IMU sensors, learning-based for the first time.

Academic Service

Reviewing

Area Chair

Outstanding Reviewer Awards

Design inspired by Jon Barron's website. Thanks for providing the source code.