EVENT DOSSIER
SimFuse3D is a new method for cross-platform 3D object detection, aimed at addressing the issue of differences in point cloud distribution caused by changes in sensor height and perspective. This method uses labeled source scanning data to repair pseudo-object geometries and generate simulated observations. By using object memory to retrieve compatible instances, simulating the alignment of point clouds and view geometries to filter out clutter, and applying Confidence-Based Multi-Stage Localization Reweighting (CMLR) to correct inconsistent predictions, all components run only during the adaptation phase, keeping the detector architecture and inference graph unchanged. In six cross-platform migration tasks, SimFuse3D outperformed Pi3DET-Net in all reported AP metrics and ranked first in almost all metrics, as well as in the comparison between nuScenes and KITTI.
SimFuse3D 方法在六次跨平台迁移任务中，在所有报告的 AP 指标上均优于 Pi3DET-Net，并在几乎所有指标及 nuScenes 到 KITTI 的对比中排名第一。该方法针对传感器高度和视角变化导致的点分布差异，利用有标签源扫描数据修复伪目标几何并生成模拟观测。具体通过对象记忆检索兼容实例、目标模拟对齐点云与视图几何并过滤杂波，以及置信度引导多阶段定位重加权（CMLR）来修正预测不一致问题。所有组件仅在适应阶段运行，保持检测器架构和推理图不变。