1 Citationer (Scopus)
206 Downloads (Pure)


In thermal video security monitoring the reliability of deployed systems rely on having varied training data that can effectively generalize and have consistent performance in the deployed context. However, for security monitoring of an outdoor environment the amount of variation introduced to the imaging system would require extensive annotated data to fully cover for training and evaluation. To this end we designed and ran a challenge to stimulate research towards alleviating the impact of concept drift on object detection performance. We used an extension of the Long-Term Thermal Imaging Dataset, composed of thermal data acquired from 14th May 2020 to 30th of April 2021, with a total of 1689 2-min clips with bounding-box annotations for 4 different categories. The data covers a wide range of different weather conditions and object densities with the goal of measuring the thermal drift over time, from the coldest day/week/month of the dataset. The challenge attracted 184 registered participants, which was considered a success from the perspective of the organizers. While participants managed to achieve higher mAP when compared to a baseline, concept drift remains a strongly impactful factor. This work describes the challenge design, the adopted dataset and obtained results, as well as discuss top-winning solutions and future directions on the topic.

TitelComputer Vision - ECCV 2022 Workshops, Proceedings
RedaktørerLeonid Karlinsky, Tomer Michaeli, Ko Nishino
Antal sider15
Publikationsdato18 jan. 2023
ISBN (Trykt)978-3-031-25071-2
ISBN (Elektronisk)978-3-031-25072-9
StatusUdgivet - 18 jan. 2023
BegivenhedEuropean Conference on Computer Vision : Workshop on Real-World Surveillance - Tel Aviv, Israel
Varighed: 23 okt. 202227 okt. 2022


WorkshopEuropean Conference on Computer Vision
ByTel Aviv
NavnLecture Notes in Computer Science

Bibliografisk note

Publisher Copyright:
© 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.


Dyk ned i forskningsemnerne om 'ChaLearn LAP Seasons in Drift Challenge: Dataset, Design and Results'. Sammen danner de et unikt fingeraftryk.