Multi-view 3D object detection has garnered significant interest due to its crucial role in autonomous driving systems. However, current research encounters challenges in detecting small and distant objects that occupy fewer pixels, as their semantic information is relatively weak and they are prone to interference from background or other object information. To address these challenges, we aim to utilize the preluded 2D (Pre2D) object detection results to facilitate multi-view 3D object detection. Specifically, we have developed a class-aware attention mechanism that utilizes class scores to enhance the feature maps with class-specific features, making foreground information more distinguishable. Furthermore, we have designed a 2D-assisted dynamic query generator that produces adaptive queries with strong alignment in both spatial and appearance, enabling the queries to locate corresponding features easily. We conducted experiments on the nuScenes dataset, and the results indicate that our approach achieves state-of-the-art performance.