Understanding Decisions of Object Detectors via Saliency Maps

The opaque nature of deep neural networks frequently prevents them from offering explanations to users. This limits their direct application in high-risk scenarios like autonomous driving and industrial control systems. Among them, object detection networks are commonly used architectures in deep neural networks, and enhancing the interpretability of object detection networks is crucial for promoting further research and practical application development. Therefore, this paper focuses on typical object detection networks, utilizing the feature statistical analysis method. By utilizing the output from the final convolutional layer of the object detection network for backpropagation, the weights of each channel of the last convolutional layer are computed using the information from backpropagation, thus obtaining an explanation for the object detection network. This approach requires less computation time and can quickly generate explanation images with low noise and smoothness.