On the 14th of December 2019 at 3pm, our Traffic4cast 2019 Core Competition concludes with a prize giving symposium at this year’s NeurIPS conference in Vancouver.
Selected leading contributors present their solutions, amongst them our top 3 ranked teams. Keynote speakers will be the IARAI‘s very own Sepp Hochreiter and Leonid Sigal from the University of British Columbia. Further presentations will explore the wider context of the competition – its relevance to traffic prediction as well as its relation to classical movie prediction tasks.
The Traffic4cast 2019 Core Competition was novel both in the size of the underlying real world data as well as in its ground-breaking approach of reformulating the problem of traffic forecasting as a movie frame prediction task. Contestants were given complete Traffic Map Movies for Berlin, Istanbul, and Moskau for 285 days. They then had to predict missing movie parts in 72 days for these 3 cities as described below. Each pixel represented an aggregation of the volume, average speed, and heading for a 100m x 100m real-world square for a 5 min interval, where the numerical values were encoded in the red-green-blue (RGB) colour values. For each city, each frame contained 495 × 436 pixels, thus contestants could train on 3 × 285 × 288 x 496 x 436 x 3 = 159,430 million values. For each city, for each of the 72 test set days, contestants were given 5 blocks of 12 consecutive images corresponding to 1 hour 5 min of real-world time, with missing data for some hours after that. Following each block, everyone then had to forecast the next 3 images, effectively predicting traffic 15 min into the future. The total number of forecast images was thus 3×5×3×72 = 3,240, of 495 x 436 RGB pixels = 647,460 values predicted. The accuracy of predictions was scored using the traditional mean squared error.
We spoke to the three Directors of IARAI to learn about the significance of this competition:
This competition is special alone because of the sheer scope and size of the data,
says Sepp Hochreiter.
We have set an important milestone in bringing real-world industrial-scale data into the academic open. We we can see from the enthusiastic response of the community that this fills a burning need. We will be delighted to extend both data scope and challenge questions jointly with the scientific community,
finds David Kreil. Michael Kopp adds
The importance of bringing together diverse fields of researcher cannot be overemphasized. The community applied the latest advances in machine learning at scale to real-world problems driven by massive data. We believe that some of humanity’s most pressing problems will only be solvable this way.
Learn more about the Traffic4cast Competition, and ongoing efforts at IARAI to spearhead and extend this approach to general geospatial processes, such as assessing climate change, weather patterns, pollution, etc.