Explaining YOLOv9 Object Detection with Attribution-Based Heatmap Methods

This study investigates the problem of explainability for visual object detectors, focusing on the YOLOv9 object detection model. The model is fine-tuned on the VOC2012 dataset using a pre-trained YOLOv9 model. The study demonstrates various attribution-based heatmap methods into the YOLOv9 architecture and analyzes their effectiveness in explaining individual detections. The attribution methods explored include Grad-CAM++, Grad-CAM, XGrad-CAM, Eigen-CAM, HiResCAM, LayerCAM, RandomCAM, and Eigen Grad-CAM. The analysis reveals that different methods provide unique perspectives about the model’s functioning, suggesting that the choice of attribution-based method can significantly influence the interpretation of the model’s predictions. In the study, Layer-CAM and HiRes-CAM consistently demonstrate superior performance in visualizing the features learned by the YOLOv9 model. This study provides valuable insights into the explainability of the YOLOv9 object detection model, highlighting the importance of selecting the right attribution method for model interpretation. The findings suggest that improving the transparency of such models can enhance their trustworthiness and applicability in real-world tasks, particularly in domains where interpretability is crucial.