Seasons in Drift: A Long-Term Thermal Imaging Dataset for Studying Concept Drift

Outdoor thermal dataset containing 8 months of video data, bounding-box annotation and weather metadata.

Ivan Adriyanov Nikolov, Mark Philip Philipsen, Anders Skaarup Johansen, Jacob Velling Dueholm, Jinsong Liu, Julio C. S. Jacques Junior, Sergio Escalera, Kamal Nasrollahi, Thomas B. Moeslund.

NeurIPS 2021

Paper | Dataset


Once computer vision algorithms step outside the lab and are deployed in real-life outdoor applications, their performance tends to drop significantly due to conditions changing over time, i.e. concept drift. Concept drift can materialize as gradual, recurring or sudden changes in the visual representation of the scene. Investigations into developing algorithms that are robust to concept drift is a widely studied topic. Traditionally the focus is put on a very specific use-case or as a method of synthetically augmenting an existing dataset. Algorithms designed in this way typically have difficulties when they are deployed outside of their specific context or see concept-drift.


We have captured 298 hours of data spanning 8 months (January 2021- May 2021) using a long wavelength infrared camera. The dataset consists of 2 minute clips at 30fps that are spaced 30 minutes apart all throughout the day. Combined with the video data each clip has associated weather data detailing weather conditions such as temperature, humidity, wind-speed, Sun-radiation intensity, etc.

Subsequently the dataset has been further extended with bounding box annotations and been used to run a object detection specific concept drift competition at the Real-World Surveillance: Applications and Challenges Workshop at ECCV2022.

The extended dataset contains 1689 clips from 188 different days spanning the time period 14/05/2020 – 30/04/2021. The extension contains 1.069.247 annotated frames with a total of 6.868.067 annotated bounding boxes and associated unique IDs allowing for tracking of each object within each clip.


Creative Commons Attribution 4.0 International


title={Seasons in Drift: A Long-Term Thermal Imaging Dataset for Studying Concept Drift},
author={Nikolov, Ivan Adriyanov and Philipsen, Mark Philip and Liu, Jinsong and Dueholm, Jacob Velling and Johansen, Anders Skaarup and Nasrollahi, Kamal and Moeslund, Thomas B},
booktitle={Thirty-fifth Conference on Neural Information Processing Systems},