BFNet: Brain-like Feedback Network for Object Detection under Severe Weather

Despite the demonstrated promising results achieved by deep learning-based object detection methods on conventional datasets, the object detection presents a significantly greater challenge when confronted with poor perception captured under adverse weather. Existing methods are studying how to improve the ability of the model to extract features by reducing the impact of severe weather on images, but these methods contradict the human visual cognition process. Based on the mechanism of visual cognition, we propose a novel object detection model based on brain-like feedback, which improves the detection performance of the model and has strong interpret ability. Specifically, an initial environment cognition assessment method is proposed, which is used to evaluate high-level prior knowledge and improve the ability to process and integrate low-level information; In addition, the amodal predictive completion method is proposed, which solves the problem of object blur and uncertainty caused by limited visual perception, and realizes the shape prediction of unknown objects. The experimental results are surprising, demonstrating the effectiveness of our method in severe weather.