Dynamic Weighting and Sparse Fusion for Robust 3D Object Detection

Currently, most mainstream multimodal 3D object detection approaches utilize dense feature representations in Bird's-Eye View (BEV). Nevertheless, in scenarios requiring long-range detection, this strategy frequently incurs substantial computational overhead and suffers from scalability constraints. To overcome these limitations, this study introduces a completely sparse multimodal framework for 3D object detection, incorporating a Dynamic-Weight Cross-Instance Interaction (DW-CIA) module and a polar coordinate-based point cloud enhancement module (Polar Sampling) to achieve efficient and accurate detection. Specifically, DW-CIA adaptively adjusts the interaction weights based on the category, distance, and size of the targets, focusing on key instances to enhance detection performance while reducing the computational burden of irrelevant instances, thus improving the model's efficiency and robustness. Meanwhile, the polar coordinate-based point cloud enhancement method supplements geometric features through target grouping and densification, substantially improving detection performance in sparse scenarios. Experiments on multiple mainstream datasets (such as nuScenes and Argoverse 2) demonstrate that this framework not only excels in sparse point clouds and complex scenes but also maintains efficient computation while achieving state-of-the-art detection accuracy, providing a new approach for long-range perception.