NightWatch-TIR-v2: Advancing Thermal Wildlife Species Detection Through Explainable Benchmarking

Abstract

Thermal Infrared (TIR) imaging has become an important sensing modality for nocturnal wildlife monitoring because it enables object perception under challenging illumination conditions where conventional RGB imaging systems fail. However, publicly available thermal wildlife datasets primarily focus on image-level classification or application-specific scenarios, providing limited support for localization-aware object detection and explainable model analysis. This paper presents NightWatch-TIR-v2, a comprehensive thermal wildlife detection benchmark comprising 10,765 manually annotated thermal images spanning eleven wildlife species, with high-quality object-level annotations provided in both YOLO and COCO formats. To establish representative performance baselines, five state-of-the-art object detection architectures, namely YOLOv8m, YOLO11m, RT-DETR-l, Faster R-CNN, and EfficientDet-D2, are systematically evaluated under a unified experimental protocol. Experimental results demonstrated excellent localization performance across all evaluated detectors, with EfficientDet-D2 achieving the strongest overall detection accuracy while YOLO11m providing an effective balance between accuracy and computational efficiency. Beyond conventional quantitative evaluation, this work introduces a comprehensive explainability framework based on multi-scale EigenCAM, where unified attention maps are generated from feature representations immediately preceding the detection heads of EfficientDet-D2 and YOLO11m. The proposed framework further introduces the Object Attention Ratio (OAR) and Background Attention Ratio (BAR) to quantitatively measure attention allocation within object regions and surrounding background, enabling dataset-level and species-wise analysis of detector interpretability. Further experimental analyses quantified the distribution of detector attention inside and outside the wildlife bounding-box regions, demonstrating that at the dataset level, both high-performing detectors allocate a substantial proportion of thresholded attention beyond the annotated object boundaries. Comparative analysis further revealed the distinct attention allocation behaviours where YOLO11m showed a relatively larger proportion of attention to the object region, as compared to EfficientDet-D2, which exhibited a broader distribution with a greater proportion of attention extending outside the object boundaries. Collectively, NightWatch-TIR-v2 established a standardized benchmark for localization-aware thermal wildlife detection and provided a unified evaluation framework that jointly assessed detection performance and model interpretability, thereby supporting future research in explainable thermal wildlife perception.