The Discovery Challenges will take place during the ECML PKDD 2023 conference. All Challenges will begin in early April 2023. Final submission and announcement of winners will be in June 2023.
Human Activity Segmentation Challenge
Human activity recognition (HAR) is essential for health monitoring, personal security, and various other applications. Current HAR systems process fixed-length subsequences of sensor data, e.g. 1 second, leading to performance losses for longer complex activities. This challenge focuses on improving accuracy of such systems by exploring the under-studied area of partitioning real-world, multi-variate human motion sensor data into variable-sized activity segments. You will work with a large new data set featuring 10.7 hours of multi-dimensional smartphone sensor data from students performing 100 different daily activities. Your objective is to accurately segment this complex motion data into an unknown number of single activities.
- Arik Ermshaus, Humboldt-Universität zu Berlin, Berlin, Germany
- Patrick Schäfer, Humboldt-Universität zu Berlin, Berlin, Germany
- Ulf Leser, Humboldt-Universität zu Berlin, Berlin, Germany
- Anthony Bagnall, University of East Anglia, Norwich, United Kingdom
- Romain Tavenard, Université de Rennes 2, Rennes, France
- Colin Leverger, Orange Innovation, Rennes, France
- Vincent Lemaire, Orange Labs, Lannion, France
- Simon Malinowski, UniversitÃ© de Rennes 1, Rennes, France
- Thomas Guyet, Inria, Villeurbanne, France
- Georgiana Ifrim, University College Dublin, Dublin, Ireland
Ariel Space Mission Data Challenge
Throughout history, people have contemplated the possibility of extraterrestrial life existing beyond our planet. Although the answer remains elusive, scientists have made significant strides in this pursuit through numerous space and ground missions. The upcoming ESA Ariel Space Mission, scheduled for launch in 2029, will scrutinize the atmosphere of over a thousand exoplanets, aiding in the generation of a comprehensive understanding of the varied nature of these remote worlds. Currently, the scientific community relies on sampling-based approaches, such as MCMC, to extract vital information from atmospheric data. However, this method is not scalable for large datasets. The Ariel Data Challenge aims to bring challenging issues faced by the Ariel space mission to the wider scientific community. This edition invites individuals and teams of all career stages to devise an innovative and scalable solution to infer crucial atmospheric properties of exoplanets from simulated Ariel observations. We warmly welcome all interested participants to join the event. Additionally, we are pleased to announce that a limited number of participants will receive free HPC support for the first time.
- Kai Hou Yip, University College London, London, United Kingdom
- Ingo Waldmann, University College London, London, United Kingdom
- Quentin Changeat, University College London, London, United Kingdom, European Space Agency (ESA), Space Telescope Science Institute (STScI), Baltimore, USA
- Mario Morvan, University College London, London, United Kingdom
- Nikolaos Nikolaou, University College London, London, United Kingdom
- Giovanna Tinetti, University College London, London, United Kingdom
ChaBuD: Change detection for Burned area Delineation
The adoption and availability of geospatial data have increased tremendously in the last decade due to the high availability of sensors, drones and aerial imagery acquisitions. Among these, remote sensing plays an important role in the geospatial industry, enabling the collection of a great amount of data at continental scale in a short amount of time. Such data represent a valuable resource to researchers and public authorities (e.g., civil protection, first responders) interested in the Earth Observation field. In particular, raster geospatial data, such as satellite acquisitions, can provide meaningful information to rescuers in case of catastrophic events, such as landslides, floodings, earthquakes and forest wildfires.
In this challenge, you will analyze bi-temporal satellite data related to California Forest Fires to predict whether an area was previously affected by the catastrophic event. All input data were collected from the Sentinel-2 mission. The developed machine learning models could support local authorities in monitoring damaged regions to plan and support restoration.
- Luca Colomba, Politecnico di Torino, Turin, Italy
- Paolo Garza, Politecnico di Torino, Turin, Italy
- Dino Ienco, INRAE, UMR TETIS, University of Montpellier, Montpellier, France
- Daniele Rege Cambrin, Politecnico di Torino, Turin, Italy
- Claudio Rossi, Senior Researcher, LINKS Foundation, Turin, Italy