Eddy Ilg

Eddy Ilg is a computer vision researcher whose work spans optical flow, uncertainty estimation and 3D reconstruction. He coauthored FlowNet in 2015 and was first author of FlowNet 2.0 in 2017; he coauthored a 2018 paper on optical-flow uncertainty. After earning a PhD at the University of Freiburg, he spent three years at Facebook (now Meta) Reality Labs on 3D object reconstruction in the wild. He joined ELLIS and took a tenure-track professorship at Saarland University in 2022, where he led the Computer Vision and Machine Perception lab’s work on 3D reconstruction and continual learning; he became a principal investigator in RTG 2853 in 2023 and a full professor at UTN in 2024, when the lab moved there and he became a NeurIPS area chair. He received a €1.5 million ERC Starting Grant in 2025 for 4DLang, which studies how language and vision can work together in AI; he is Professor of Computer Vision at UTN, and his March 2026 talk explores how 3D reconstruction can link spatial understanding, language and world models.

Now at: Saarland University and European Laboratory for Learning and Intelligent Systems

Timeline

2015

FlowNet learns optical flow

In 2015, he coauthors FlowNet, which frames optical-flow estimation as supervised learning with convolutional networks. The authors train on the synthetic Flying Chairs dataset; their networks generalize to Sintel and KITTI at 5 to 10 frames per second. The paper appears at ICCV in December 2015.
2017

FlowNet 2.0 cuts error

In 2017, he is first author of FlowNet 2.0, which stacks networks and adds a subnetwork for small motions. The system cuts the original FlowNet’s estimation error by more than 50% and runs at interactive frame rates.
2018

Uncertainty in optical flow

In 2018, he coauthors a network that estimates local uncertainty in optical-flow predictions. The architecture uses a Winner-Takes-All loss to produce complementary hypotheses and uncertainty estimates in one forward pass.
By 2022

Three years at Reality Labs

Before he joins Saarland University in 2022, he completes a PhD at the University of Freiburg and spends three years at Facebook (now Meta) Reality Labs, focusing on 3D object reconstruction in the wild.
2022–2023

Saarland professorship and ELLIS faculty

He joins ELLIS in May 2022 and takes a tenure-track professorship at Saarland University in June 2022. He leads the Computer Vision and Machine Perception (CVMP) lab, which combines 3D reconstruction with continual learning to build machine perception algorithms that adapt to their environment and evolve. He recruits for 3D computer vision and continual learning, and becomes ELLIS Unit Faculty in November 2022. In May 2023, the DFG accepts RTG 2853, Neuroexplicit Models of Language, Vision and Action, with him as one of its principal investigators.
2024

CVMP moves to Nuremberg

In March 2024, a CVMP lab paper on Neural Point Cloud Diffusion for Disentangled 3D Shape and Appearance Generation is accepted to CVPR. In May 2024, he becomes a NeurIPS area chair; in August 2024, he receives a full professorship at UTN, and the lab moves there in October 2024. His earlier work includes optical-flow publications and development of AR technologies for Meta.
September 2025

An ERC Starting Grant

In September 2025, he receives a European Research Council Starting Grant for 4DLang, a four-year project funded with €1.5 million to explore how language and vision can be combined in AI.
March 2026

Professor at UTN

In March 2026, he is Professor of Computer Vision at UTN. His talk presents the team’s spatio-temporal language, which uses 3D reconstruction to connect spatial perception with natural-language reasoning and enable world models.