Physically Interpretable Object Detection in Foggy Point Cloud
推荐理由
Under ideal conditions, 3D object detection (3DOD) provides accurate spatial position and object classification information. However, in adverse weather condit…
核心判断
论文摘要（中文）
论文提出 IFOD 雾天点云目标检测架构，以大气散射模型提供物理先验，包含 Focus Sparse Convolution（FSC）和 ASM-Guided Enhancement（AGE）；同时构建由物理散射模型模拟的 FogKITTI 数据集以缓解标注不足。
Under ideal conditions, 3D object detection (3DOD) provides accurate spatial position and object classification information. However, in adverse weather conditions such as foggy, LiDAR signals suffer from severe scattering, resulting in abnormal point cloud density and distance shifts. This significantly affects the accuracy and reliability of 3DOD. Existing methods in foggy conditions are mainly limited by two factors: first, the scarcity of annotated point cloud data in foggy weather; and second, the lack of foggy-aware guidance mechanisms in existing deep learning frameworks. To address these issues, we propose physically Interpretable Object Detection in Foggy Point Cloud (IFOD), a novel architecture that employs physically-motivated design principles through an Atmospheric Scattering Model (ASM) to achieve robust perception in foggy conditions. The IFOD architecture includes two key innovative components: the Focus Sparse Convolution (FSC) module and the ASM-Guided Enhancement (AGE) module. FSC improves information sparsity issues by adapting the position of output elements. AGE, in addition to maintaining the original 3D spatial branch, generates a 2D pseudo-image feature space branch to extract key fog features for encoding 3D features, and this consistency checking mechanism helps to reduce false predictions and improve the model’s ability to recognize object in foggy conditions. Furthermore, to address the data labelling scarcity problem, we introduce FogKITTI, an extended dataset simulated through physical scattering models, providing valuable resources for 3D point cloud perception research in foggy environments. Experimental results show that IFOD achieves superior detection performance on both realworld and simulated datasets.
研究动机（中文总结）
雾天激光雷达会因散射出现点云密度异常和距离偏移，降低 3D 检测可靠性；现有方法还受雾天标注数据稀缺、缺少雾感知机制的限制。
Furthermore, to address the data labelling scarcity problem, we introduce FogKITTI, an extended dataset simulated through physical scattering models, providing valuable resources for 3D point cloud perception research in foggy environments.
创新与贡献（中文总结）
把大气散射物理模型融入点云检测，并同时提出 FSC、AGE 两个结构模块和 FogKITTI 扩展数据集。
To address these issues, we propose physically Interpretable Object Detection in Foggy Point Cloud (IFOD), a novel architecture that employs physically-motivated design principles through an Atmospheric Scattering Model (ASM) to achieve robust perception in foggy conditions.
方法与证据
方法（中文总结）
FSC 调整输出点位置以缓解点云稀疏问题；AGE 保留原有 3D 分支，并增加 2D 伪图像雾特征分支来增强 3D 表示，通过一致性检查减少误检。
To address these issues, we propose physically Interpretable Object Detection in Foggy Point Cloud (IFOD), a novel architecture that employs physically-motivated design principles through an Atmospheric Scattering Model (ASM) to achieve robust perception in foggy conditions.
实验结果（中文总结）
摘要称 IFOD 在真实和模拟数据集上均取得更好的检测表现；当前提取的结果依据没有给出数据集名称、指标数值和基线差异，需回看实验表格补齐。
Experimental results show that IFOD achieves superior detection performance on both realworld and simulated datasets.