DMS3D: Multi-Modal 3D Object Detection Algorithm Based on Discriminant Fusion Strategy and Multi-Scale Feature Fusion

Abstract

Multi-modal 3D object detection, which integrates point clouds and images, has gained increasing attention due to its potential to enhance detection accuracy. However, challenges such as redundant information and scale inconsistencies hinder effective feature fusion. To address these issues, we propose DMS3D (Multi-modal 3D object detection algorithm based on discriminant fusion strategy and multi-scale feature fusion), a Discriminative fusion strategy and Multi-scale feature fusion-based System for 3D object detection. Our method introduces two key modules: (1) a multi-scale feature fusion module (MSFF) based on graph neural networks, which refines point cloud representations by aligning them with image semantics while preserving fine-grained details; and (2) a discriminative fusion module (DFM) with a lightweight gating mechanism, which selectively fuses multi-modal features to reduce redundancy and noise while enhancing computational efficiency. Extensive experiments on the KITTI and nuScenes datasets validate the effectiveness of DMS3D, demonstrating superior detection performance compared to state-of-the-art methods.