Traffic movie frames

Traffic Movie frames. Each frame covers urban area of ~ 50km x 50km.
Pixels are ~ 100m x 100m aggregating GPS probe volumes and speeds into 5 minute intervals.

The Data

Our unique industrial-scale dataset provides a high resolution privacy preserving view of urban traffic. The dataset covers 10 culturally diverse cities around the world in a time span of 2 years and includes dynamic and static information. The dynamic data are derived from the GPS trajectories of a large fleet of probe vehicles, made available by HERE Technologies. The GPS trajectories are aggregated into spatiotemporal cells. Each cell corresponds to the area of approximately 100m x 100m and the time interval of 5 minutes. The dynamic information is encoded in 8 channels containing the traffic volume and average speed per heading direction: NE, SE, SW, or NW. The static information describes the properties of the road maps and is split only into spatial cells. It is encoded in 9 channels representing the density of the road network and the road connections to the 8 neighboring cells. The traffic data can be presented as a movie with 288 frames per day for each city, thus effectively recasting traffic prediction as a video frame prediction task.

Spatial tessellation of the road network.

Spatial tessellation of the road network. 
Spatial cells with GPS probes (left) and their magnification (two boxes right)
showing complexity of the road network.

Dynamic data

Dynamic data: 8 channels encode traffic volume and average speed per direction.

Core and Extended Challenge

Going beyond the challenges of the past competitions, Traffic4cast 2021 focuses on models that adapt to domain shifts in space and time. For training, we provide dynamic data for 4 cities in the effective format introduced in Traffic4cast 2020. The training data cover 6 months from 2019 and 6 months from 2020 during the COVID-19 pandemic. These dynamic data are complemented by static information on road geometry. Traffic4cast 2021 comprises two transfer learning challenges:

  • In the core challenge, participants are tasked to handle temporal domain shift in traffic due to COVID-19. In addition to the training data described above, we provide 6 months of training data from 2019 plus 100 1-hour test slots from 2020 – for 4 different cities. The challenge is to predict the dynamic traffic states 5, 10, 15, 30, 45 and 60 minutes into the future after each of the test slots for each of these additional 4 cities.

  • In an extended challenge, participants are asked to predict dynamic traffic states for 2 new cities, hitherto entirely unseen. As in the core challenge, traffic needs to be predicted for each city 5, 10, 15, 30, 45 and 60 minutes into the future following 100 1-hour test slots. The test slots and the static maps are the only data provided for these cities. Moreover, these test slots are randomly chosen: 50 from 2019 and 50 from 2020 – without revealing the year!

A solution to the extended challenge could also be applied to the core challenge, thus truly making it more universal. Both challenges are few-shot learning challenges, an actively developing area of AI, and can be tackled in a multitude of ways given the data provided.

The common underlying scientific question is how to build robust models that can predict how a complex spatial process evolves over time, so that the models can swiftly adjust to domain shifts both in space and time after seeing only a few examples from the new regime.

Temporal domain shift - Istanbul

Illustration of temporal domain shift due to pandemic. Traffic volume (sum of all directions)
in a 24h interval in 2019 (blue) and 2020 (orange) in Istanbul (top) and New York (bottom).

Prizes Core and Extended Challenge

The winners of the core challenge (temporal domain shift) will receive the following prizes:

1st place – a voucher or cash prize worth 8,000 EUR to the participant/team plus one free NeurIPS 2021 conference registration;

2nd place – a voucher or cash prize worth 4,000 EUR to the participant/team plus one free NeurIPS 2021 conference registration;

3rd place – a voucher or cash prize worth 2,000 EUR to the participant/team plus one free NeurIPS 2021 conference registration.

The winners of the extended challenge (spatiotemporal domain shift) will receive the following prizes:

1st place – a voucher or cash prize worth 8,000 EUR to the participant/team plus one free NeurIPS 2021 conference registration;

2nd place – a voucher or cash prize worth 4,000 EUR to the participant/team plus one free NeurIPS 2021 conference registration;

3rd place – a voucher or cash prize worth 2,000 EUR to the participant/team plus one free NeurIPS 2021 conference registration.

To collect a prize, participants need to adhere to the competition’s T&Cs. These require the publication of working solutions (including learnt parameters) as well as a short scientific paper describing their approach on GitHub and arXiv, respectively.

Timeline Core and Extended Challenge

May 28, 2021: Traffic4cast competition forums open.

June 15, 2021: Traffic4cast competition goes live. Traffic data are available for download.

October 14, 2021, 23:59 CET: Submission deadline.

October 21, 2021, 23:59 CET: Extended Paper and Slides Submission, publication of code deadline, start of paper review

November 5, 2021: Announcement of awards.

December 6–14, 2021: NeurIPS Conference.

Analysis and Interpretability Challenge

The core and extended challenge in Traffic4cast 2021 aim at finding models that adapt to domain shifts in space and time. It’s an interesting observation to see that, also within the new transfer learning setting, UNet architecture based solutions still dominate the leaderboard when combined with a range of new and creative approaches utilizing ensembles, domain adaptation, patch-based subdivision and the underlying road graph.

The winning approaches show that the transfer across temporal as well as spatial domain shifts is possible. In the core competition which challenged participants to predict only the temporal shift, for instance, and with training data available for all cities, no city-specific models seemed to be needed. It’s therefore plausible that a single model could be trained to make predictions for all cities, evidently a tantalizing prospect for application scenarios predicting traffic on a global scale.

The solutions by the participants also help to better understand the limitations and open questions concerning the data and prediction tasks, such as how the road graph aspects could be formulated so that graph neural network approaches become competitive. The handling of sparse input data, for example, at night or in less dense areas, is another topical area of future research. This is also related to the more general problem of choosing an appropriate metric. Standard MSE often leads to insufficient modelling of sparse data, instead favoring predictions that regress to the mean. 

These questions are of interest to the wider academic community and, indeed, to anyone researching traffic predictions at scale. Hence, the Analysis and Interpretability Challenge addresses these questions as an interdisciplinary challenge on ML in Traffic Engineering. Specifically we invite colleagues world-wide to submit their latest research addressing any of the following key topical focus areas:

  • Analysis and Interpretability of traffic predictions:

    • What does the model learn? E.g. rules vs statistics.

    • Comparison and analysis of competition winning models with other benchmarks or potentially new models.

    • How do the models deal with outliers and anomalies?

  • Traffic representations and data

    • Understanding the data, e.g. temporal properties, correlations of volume and speed

    • Dealing with sparsity and unknown probe distributions

    • Representing and using enhanced road graphs

    • Visualizations for understanding traffic data and predictions

  • Task formulation and metrics

    • Understanding the loss function

    • From generic MSE to distributional, traffic and/or physics aware metrics

    • Graph-based metrics

All submitted work will be peer-reviewed by the Scientific Committee of the workshop, and we encourage both quantitative as well as qualitative analyses. Specifically, prices are awarded by the Scientific Committee and are solely based on academic merit.

Data

The data is the same as for our core/extended competition – please find instructions on how to download in the competition forum. Like the core and extended challenge this challenge is governed by the competition T&Cs, in particular with regards to the access and use of the data as well as the publication requirements.

In addition we’re also providing the submitted predictions from the Core and Extended Challenge. Please find more information on how to obtain this data in this forum post.

Submission

Please submit your work via EasyChair (https://easychair.org/my/conference?conf=t4c2021#) and give us a heads-up about your submission to traffic4cast@iarai.ac.at, we will have an updated index of all submissions on this page. Feel free to also publish a preview of your work via arXiv or other channels. Corresponding code and other results should preferably be published via GitHub and linked to from the papers.

Prizes

Invitation to the Traffic4cast workshop in Vienna, June 2022, for accepted papers (based on submissions reviews)

Special price (best paper award): 1 year IARAI fellowship

Timeline

December 10, 2021: challenge kick-off at IARAI Traffic4cast summit, submitted predictions from the Core Challenge and Extended Challenge are available for download

May 31, 2022: submission deadline for papers, code and analysis results

June 10, 2022: invitations to the Traffic4cast Workshop in Vienna

July 5-7, 2022: announcement of awards during the Traffic4cast Workshop

©2023 IARAI - INSTITUTE OF ADVANCED RESEARCH IN ARTIFICIAL INTELLIGENCE

Imprint | Terms and conditions

Log in with your credentials

or    

Forgot your details?

Create Account