Enhancing Object Detection Robustness: A Synthetic and Natural Perturbation Approach

This paper addresses the robustness of object detection models against real-world distribution shifts, focusing on natural perturbations like varying lighting, blur, and brightness. Four state-of-the-art models (Detr-ResNet-101, Detr-ResNet-50, YOLOv4, and YOLOv4-tiny) were evaluated using the COCO 2017 and ExDark datasets. Synthetic perturbations, simulated using the AugLy package, were used to enhance model robustness through data augmentation. An ablation study assessed the impact of synthetic perturbations on performance against real-world shifts, revealing a tangible link between synthetic augmentation and real-world robustness. The findings highlight the effectiveness of synthetic perturbations in improving model robustness and offer insights for developing more reliable object detection models.

Nilantha Premakumara, Brian Jalaian, Niranjan Suri, Hooman Samani