Meta-token Learning for Thermal Object Detection with Transformers

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