Adaptation of object detection networks under anomalous conditions

Abstract

Out of distribution (OOD) samples can negatively affect the performance of deep neural networks. When deep neural networks are used in cyber-physical systems, it may be vulnerable to OOD data, leading to large errors and compromise the safety of the system. This paper proposes combining OOD explanation and object detection model into a pipeline that maximizes efficiency and maintains accuracy across different OOD situations. We use YOLOv7 as the object detection model, which accepts different input sizes with a single set of weights. Under anomalous conditions, we can increase the input size to reduce the drop in accuracy at the expense of speed. This allows accuracy and speed to be balanced under different anomalous conditions. Alternatively, fine-tuned weights can be switched in under anomalous conditions, which shows consistent improvements though at higher costs.