A controlled ablation study of CBAM in YOLOv8 and YOLOv9 for multi class weapon detection

Abstract

Recent YOLO-based object detectors provide a strong accuracy–latency trade-off for security-critical applications, yet it remains unclear whether attention mechanisms consistently improve modern architectures. This paper presents a controlled ablation study of integrating the Convolutional Block Attention Module (CBAM) into YOLOv8 and YOLOv9 for multi-class surveillance object detection. Experimental results reveal a clear architecture-dependent behavior. CBAM improves YOLOv8 performance, increasing precision from 0.7786 to 0.8060 and mAP@50 from 0.8571 to 0.8753 with modest computational overhead. In contrast, CBAM significantly degrades YOLOv9 performance, reducing mAP@50 from 0.9501 to 0.7600 and mAP@50–95 to 0.5400, while nearly doubling the model size (25.3–48.7 M parameters). These findings demonstrate that attention mechanisms are not universally beneficial; rather, their effectiveness depends on the underlying architecture. In high-capacity models such as YOLOv9, CBAM introduces redundancy that negatively impacts performance. Additional analysis, including efficiency evaluation and Grad-CAM visualization, further supports this interpretation. Overall, this study highlights that attention should be treated as an architecture-dependent design choice rather than a generic performance enhancement, providing practical insights for the development of efficient real-time detection systems.