VIULab
Vision and Information Understanding Laboratory (VIULab)
University of Liège
Embedded Files
6
6
CVPR/ECCV papers since 2024
30+
30+
Sports-AI publication
5
5
CVsports best paper awards
9
9
Researchers
Welcome to the Vision and Information Understanding Laboratory (VIULab) at the University of Liège (ULiège)!
VIULab is the computer vision and AI research lab of the University of Liège. We build models that understand the visual world in motion, following the action in hours of video, reconstructing scenes in 3D, and rethinking how vision models learn and how we evaluate them. Our research appears at the top venues in the field such as CVPR and ECCV, and reaches real-world applications in sports analytics, remote sensing, and autonomous driving, including SoccerNet, the largest open benchmark for soccer video understanding, which we co-organize. Explore our latest work andcome build the next generation of video AI with us.
Highlighted works!
Highlighted works!
MeshSplatting: Differentiable Rendering with Opaque Meshes - by Jan Held @CVPR2026 (oral)
MeshSplatting: Differentiable Rendering with Opaque Meshes - by Jan Held @CVPR2026 (oral)
Gaussian Splatting looks stunning but can't run in the mesh-based engines that power games and AR/VR. MeshSplatting reconstructs real, opaque meshes directly through differentiable rendering, beating the state of the art by +0.69 dB while training 2× faster on half the memory.
What Is the Optimal Ranking Score Between Precision and Recall? - by Adrien Deliège @CVPR2026
What Is the Optimal Ranking Score Between Precision and Recall? - by Adrien Deliège @CVPR2026
For fifty years, the F-score has been the default compromise between precision and recall, but is it actually a good one? We prove F1 is far from optimal and give a closed-form way to find the best trade-off for any benchmark.
TrackMAE: Video Representation Learning via Track Mask and Predict - by Renaud Vandeghen @CVPR2026
TrackMAE: Video Representation Learning via Track Mask and Predict - by Renaud Vandeghen @CVPR2026
Masked video models learn what things look like, but not how they move. TrackMAE feeds point-tracking trajectories into the reconstruction signal, beating state-of-the-art video self-supervised learning across six datasets.
We co-organize SoccerNet, the largest open benchmark for soccer video understanding. Open datasets, annual challenges at CVPR, and a global community of researchers and sports-tech teams pushing the state of the art together. → soccer-net.org · Join our Discord
VIULab · Montefiore Institute, University of Liège, Belgium
Page updated
Google Sites
Report abuse