Knowledge-Aware Object Detection Adjusting mAP for Class Hierarchies in Autonomous Vehicles

Traditional object detection evaluation metrics often fail to account for the hierarchical nature of object classifications, leading to rigid assessments where semantically similar misclassifications are treated as entirely incorrect. This limitation can be particularly problematic in autonomous vehicle applications, where understanding subclass-superclass relationships is crucial for safe decision-making. This study introduces a hierarchical evaluation framework that assigns partial credit to predictions based on class relationships, offering a more flexible and informative assessment of object detection models. Our framework uses the YOLOv5 object detection model trained on a synthetic dataset generated with CARLA, an autonomous driving simulator. By leveraging synthetic data with hierarchical labels, we examine how integrating semantic relationships into evaluation metrics impacts model performance. Experimental results highlight the potential benefits of hierarchical evaluation, demonstrating improved interpretability of detection performance and better alignment with real-world decision-making processes in autonomous driving. This work underscores the importance of incorporating semantic awareness into object detection assessments and the role of synthetic datasets in advancing hierarchical classification models.