Layer-wise Approach to Activation Analysis in YOLOv8 Model: Measuring Localization Effectiveness through GradCAM

This research analyzes activation patterns across YOLOv8's three scale-specific convolutional layers using GradCAM. We measure activation density and top-k pixel distribution to quantify detection effectiveness for small, medium, and large objects. Our novel scale classification model determines object size based on layer-specific activation patterns, providing insights into YOLO's multi-scale detection mechanisms