Pascal Mettes
Assistant Professor - University of Amsterdam
Computer vision
Hierarchical knowledge
Hyperbolic geometry
My research goal is to find a bridge between prior knowledge and deep networks. Deep learning in computer vision thrives under examples but commonly ignores additional explicit knowledge about the research problem. Think about hierarchical relations between categories, relational knowledge between tasks, or spatio-temporal knowledge. The research of my team focuses on discovering the shared geometry between pixels and knowledge. Specifically, we research hyperbolic and hyperspherical geometry for deep learning and learning with prototypes. Our advances and new research proposals are investigated for computer vision problems such as hierarchical recognition, segmentation, localization, and zero-shot recognition in images and videos.
[08-2023] "Poincaré ResNet" accepted to ICCV 2023.
[08-2023] "HypLL: The Hyperbolic Learning Library" is accepted to ACM Multimedia 2023.
[06-2023] HypLL: The Hyperbolic Learning Library is now available! [github] [paper]
[06-2023] We are organizing a new tutorial on hyperbolic learning at CVPR 2023.
[06-2023] I am giving a keynote at the CVPR 2023 anti-UAV workshop.
[05-2023] Our survey on Hyperbolic Deep Learning in Computer Vision is available online.
[02-2023] The recordings of the ECCV 2022 hyperbolic tutorial is now available online.
[11-2022] "Maximum Class Separation as Inductive Bias in One Matrix" accepted to NeurIPS 2022.
[10-2022] We are organizing a tutorial on Hyperbolic Representation Learning at ECCV 2022.
[08-2022] "Less than Few: Self-Shot Video Instance Segmentation" accepted to ECCV 2022.
[09-2022] We are organizing the second Video Symposium in Amsterdam.
[06-2022] "Hyperbolic Image Segmentation" accepted to CVPR 2022.
[04-2022] We are organizing the Netherlands Conference on Computer Vision 2022.
Host: prof. Zeynep Akata
Host: prof. Shih-Fu Chang
Promotors: prof. Cees Snoek and prof. Arnold Smeulders
Code and pre-computed prototypes are available [here].
Pre-trained models can be downloaded by following the instructions [here].
Code and pre-computed prototypes are available [here].
The complete Video Water Dataset is available [here] (warning 14GB).