Enhancing Resilience: Redundancy Reduction for Object Detection in Adversarial Conditions

Vision systems, like other deep learning-based systems, encounter limitations due to training data and struggle to handle adversarial conditions such as varying lighting and weather conditions. In this paper, we propose a knowledge distillation framework aimed at bolstering the resilience of computer vision systems under adversarial conditions. Specifically, we focus on object detection task in adverse weather conditions and demonstrate that our system either exceeds or matches the state-of-the-art accuracy levels. Our system achieves a 2% higher mean average precision (mAP@50) in hazy conditions, and 9% higher mean average precision (mAP@50) in low-light conditions, compared to the nearest state-of-the-art frameworks.