
RWS introduces a new challenge on Robust Thermal-Image Object Detection using the LTDv2 dataset, benchmarking multi-object detection under long-term thermal drift. Real-world conditions like ambient changes, sensor aging, and weather dynamics cause performance degradation, especially in thermal imagery, which relies on shape cues and temperature differentials.
LTDv2 enables quantitative assessment of model robustness to appearance shifts. Submissions will be evaluated on overall mAP and temporal consistency. A development phase for method validation and a test
phase with submission limits to avoid overfitting are included. Baseline YOLOv8 detectors will be provided.
This challenge is organized by Anders S. Johansen, Andreas Aakerberg, and Kamal Nasrollahi (Aalborg University, Denmark), together with Marco Parola and Mario G.C.A. Cimino (University of Pisa, Italy).
The challenge will be hosted on CodaBench, starting October, 2025. More info will appear soon…
