<p>Object detection is a vital area of computer vision that has achieved significant advancements in recent years. A common approach to detection involves refining the backbone of the model using supervision from classification tasks during the pre-training phase. However, the differences between classification and detection tasks during the pre-training and fine-tuning stages lead to challenges such as attention bias and a lack of location information. To address these issues, we introduce a lightweight and generalizable module called SegHead in this study. This module leverages segmentation feedback to align the pre-training and fine-tuning phases of object detectors, thereby optimizing the backbone parameters. Notably, our technique relies solely on bounding box-level supervision to achieve weakly supervised semantic segmentation without incurring additional costs for the detection model. We have employed a method called the Combination of GrabCut and Filling Rate (CGFR) to enhance the accuracy of segmentation supervision and generate improved pseudo masks. This approach boosts both detection accuracy and segmentation performance. Using the COCO evaluation metric, our method surpasses the RetinaNet baseline by 2.5 points in Average Precision (AP) on the PASCAL test set. Additionally, we have demonstrated in further experiments that our method can significantly improve the performance of various object detectors with minimal cost.</p>