RF-BoxCAM: Receptive Field Guided High Resolution Visual Explanations for Object Detection Networks

Kurzfassung

Aerial object detection is a fundamental component of safety-critical applications such as surveillance, traffic monitoring, and autonomous systems. Although deep learning (DL) based object detectors have achieved strong performance, their lack of interpretability remains a major barrier to trustworthy deployment. This limitation is particularly pronounced in aerial imagery, where objects are often small, densely packed, and visually ambiguous. Existing explainable artificial intelligence (XAI) techniques, both black and white-boxes, often fail to provide meaningful and faithful explanations for small-object detections, as their attribution mechanisms tend to focus on coarse or irrelevant regions. In this paper, we analyze these shortcomings and propose Receptive Field guided Bounding Box Class Activation Map (RF-BoxCAM), a novel XAI method that is specifically designed and adapted for commonly used YOLO object detection architectures and specifically designed to address the challenges of small-object detection. The proposed approach produces efficient, object-centric visual explanations in a single forward pass, improving transparency while significantly reducing computational overhead compared to black-box methods. By enhancing interpretability for small-object aerial detection, this work contributes toward safer and more reliable deployment of DL models in safety-sensitive systems.