Meta-token Learning for Thermal Object Detection with Transformers

Transformers have emerged as a dominant architecture in computer vision, demonstrating strong performance across a range of tasks. In this work, we investigate their effectiveness for object detection in thermal imagery, a setting often challenged by variable environmental conditions. We focus on a long-term thermal surveillance dataset captured using a stationary thermal camera subjected to diverse weather scenarios. To address the influence of such external factors, we propose an analysis that integrates weather information as metatokens alongside thermal images, enabling the model to account for environmental context during detection. Among the fusion strategies explored, we find that a token-level concatenation allows transformers to partially exploit the auxiliary weather data, leading to improved detection performance. Our study highlights the potential of meta-token transformer-based architectures for robust detection in challenging thermal environments.