This session features an in-depth discussion of the Traffic4cast 2021 competition results with the core and extended challenge winners, the winner of the special prize as well as selected participants from the leaderboard. The goal of this year’s competition was to build traffic models which are robust to domain shifts in space and time.
Traffic4cast was selected again for the competition track at Neural Information Processing Systems (NeurIPS), the leading event in machine learning. The summary of the competition is presented at the NeurIPS competition track on Thursday, December 9.
Related resources:
Program
Title | Presenter | Affiliation | Duration | Abstract |
|---|---|---|---|---|
1st session: Fri 2021-12-10 10:00-12:30 CET session chair: Michael Kopp |
| |||
An Introduction to the Traffic4cast Challenge | Sepp Hochreiter | IARAI, Johannes Kepler University | 15 min | |
Keynote talk: Physics, Molecule and PDE Modeling using Graph Neural Networks | Johannes Brandstetter | AMLab Amsterdam, Johannes Kepler University | 45 min | |
| The Traffic4cast Competition Design and Data | Christian Eichenberger | IARAI | 10 min | |
2nd prize core & 1st prize extended: Applying UNet for the traffic map prediction across different time and space | Sungbin Choi | - | 15 min | |
3rd prize core: Solving Traffic4Cast Competition with U-Net and Temporal Domain Adaptation | Vsevolod Konyakhin, Nina Lukashina and Aleksei Shpilman | ITMO University, JetBrains Research, HSE University | 15 min | |
3rd prize extended: Traffic Forecasting on Traffic Movie Snippets | Nina Wiedemann and Martin Raubal | MIE Lab, ETH Zurich | 15 min | |
Submission: Large-scale Traffic Prediction using 3DResNet and Sparse-UNet | Bo Wang, Reza Mohajerpoor, Chen Cai, Inhi Kim and Hai Vu | Monash University, CSIRO's Data61, Kongju National University | 15 min | |
Special prize | Moritz Neun and Christian Eichenberger | IARAI | 15 min | |
2nd session: Fri 2021-12-10 13:30-16:00 CET session chair: David Kreil | ||||
Keynote talk: Automatic Analysis of Mobility Data | Monika Sester | IKG, University of Hannover | 45 min | |
1st prize core & 2nd extended: Learning to Transfer for Traffic Forecasting via Multi-task Learning | Yichao Lu | Layer 6 AI | 15 min | |
Submission: A Graph-based U-Net Model for Predicting Traffic in unseen Cities | Luca Hermes, Andrew Melnik, Riza Velioglu, Markus Vieth and Malte Schilling | University of Bielefeld | 15 min | |
Submission: SwinUnet3D - A Hierarchical Architecture for Deep Traffic Prediction using Shifted Window Transformers | Alabi Bojesomo, Hasan Al-Marzouqi and Panos Liatsis | Khalifa University | 15 min | |
Submission: Dual Encoding U-Net for Spatio-Temporal Domain Shift Frame Prediction | Jay Santokhi, Dylan Hillier, Yiming Yang, Joned Sarwar, Anna Jordan and Emil Hewage | Alchera Data Technologies | 15 min | |
Summary, outlook & discussion | Moritz Neun and Christian Eichenberger | IARAI | 45 min |
Keynote Speakers
Keynote talk:
JKU Link, Austria / Amsterdam Machine Learning Lab, Netherlands
Johannes Brandstetter is a Senior Researcher with a history of working and publishing in Machine Learning, and High Energy Physics. During his PhD, Johannes was involved at the CMS experiment at the Large Hadron Collider at CERN, working on Higgs boson physics. Currently, Johannes is a Tenure Track Professor at the Johannes Kepler University in Linz (group headed by Prof. Sepp Hochreiter) and a postdoctoral ELLIS researcher at the Amsterdam Machine Learning Lab (group headed by Max Welling). His research comprises modeling of physics processes and simulations, developing general and fully neural solvers for Partial Differential Equations, and modeling molecular properties and dynamics.
Physics, Molecule and PDE Modeling using Graph Neural Networks
This talk spans a bridge between the modeling of simulated physical processes, equivariant graph neural networks for molecular prediction, and graph neural network based PDE solvers. These topics share many common challenges, examples of which are how to apply an accurate noise modeling, and how to achieve stability for many unrolled prediction steps. Eventually, it will be surprising how closely related these topics are. Connections to traffic data are provided in the end.
Keynote talk:
Leibniz University Hannover, Germany
Monika Sester is a surveying engineer by training (University Karlsruhe) and professor and head of the Institute of Cartography and Geoinformatics (ikg) at Leibniz University Hannover. Her projects are funded by the German Science Foundation (DFG), German Ministries, EU, as well as collaborations with Mapping Agencies and industry. She chaired several Working Groups in ISPRS (International Society of Photogrammetry and Remote Sensing) and was Vice President of the International Cartographic Association (2015-2019). She has been Vice President of Leibniz University Hannover (2016-2017). Currently, Monika is chair of a Senate Commission of the DFG on Earth System Science and member of the Senate of the Helmholtz Association. She published 138 papers in peer-reviewed international journals and conferences.
Automatic Analysis of Mobility Data
Today’s sensor technologies enable very dense detection of a wide range of dynamic environmental information: GNSS in consumer grade and professional devices lead to trajectories; mobile mapping systems scan the environment with high resolution imagery and 3D point clouds. This data allows for a variety of analyses, e.g. the prediction of traffic participants, or the determination of the dynamics of the environment. In the talk, several examples of research at the Institute of Cartography and Geoinformatics are presented.