AF-Center：基于自适应体素绘画融合和高斯中心样本分配的多模态三维目标检测

相机图像和激光雷达点云可以为3D目标检测提供互补信息，但如何进行有效的融合仍是一个挑战。针对传统方法中无区分性融合带来的对齐偏差问题，提出一个自适应融合网络。首先构建点云体素与对应的多个图像像素之间的注意力亲和矩阵，然后依据亲和矩阵实现多像素到单体素的重要性区分融合。除此之外，针对传统anchor-based检测方法难以枚举所有方向的问题，将目标表示为关键点，首先进行中心点定位，然后回归到3D尺寸与方向等其他属性。同时，针对关键点检测时中心点样本量过少的问题，使用椭圆高斯热图进行了中心点样本的再分配。该算法在Waymo数据集上，较基线PointPillar、CenterPoint与3D-MAN分别提升了2.3%、5.9%与4.0% level2 mAPH。

AF-Center: multi-modal 3D object detection method with adaptive voxel-painting fusion and Gaussian center sample assignment

Camera images and LiDAR point clouds can provide complementary information for 3D object detection, but how to perform effective fusion remains a challenge. To address the alignment bias problem caused by the undifferentiated fusion in traditional methods, this paper proposed an adaptive fusion network. Firstly, it constructed the attention affinity matrix between the point cloud voxels and the corresponding multiple image pixels, and then achieved the importance differentiation fusion from multiple pixels to single voxels based on the affinity matrix. In addition, for the problem that it was difficult to enumerate all directions in the traditional anchor-based detection method, this paper represented the target as key points and first performs center localization, and then regressioned to other attributes such as 3D size and orientation. At the same time, to address the problem that the sample size of the center was too small for key point detection, this paper used an elliptical Gaussian heat map for the redistribution of center samples. This algorithm improves 2.3%, 5.9%, and 4.0% level2 mAPH over the baseline PointPillar, CenterPoint, and 3D-MAN, respectively, on the Waymo dataset.