Robust 2D object detection is a cornerstone of perception systems in autonomous vehicles (AVs). While Convolutional Neural Networks (CNNs) like the YOLO series have long dominated real-time detection, the emergence of Vision Transformers (ViTs) introduces new paradigms in handling global context and occlusions. This paper presents a comprehensive comparative analysis of two state-of-the-art architectures: YOLOv8 (representing the peak of efficient CNN design) and RT-DETR-L (a real-time detection transformer). We evaluate these models on the extensive Waymo Open Dataset, focusing on critical safety metrics including Mean Average Precision (mAP) at varying Intersection over Union (IoU) thresholds (0.5, 0.5:0.95) and object scales (small, medium, large). Furthermore, we conduct a fine-grained robustness analysis across diverse environmental conditions, specifically varying weather (fair vs. rain) and time-of-day (day vs. night). Our methodology provides a rigorous framework for understanding the trade-offs between the inductive bias of CNNs and the global attention mechanisms of Transformers in safety-critical driving scenarios.
Karadurak et al. (Fri,) studied this question.