MD3D: Mixture-Density-Based 3D Object Detection in Point Clouds

The design factors of anchor boxes, such as shape, placement, and target assignment policy, greatly influence the performance latency 3D object detectors. Unlike image-based 2D anchors, anchors must be placed in a space determined differently for each class different sizes. This imposes significant burden on complexity. To tackle this issue, various studies have been conducted how to set form. However, practical reasons, anchor-based methods select by compromising between latency. Consequently, only objects that are similar shape size an can obtain high accuracy. In paper, we propose Mixture-Density-based Object Detection (MD3D) point clouds predict distribution bounding boxes using Gaussian Mixture Model (GMM). With anchor-free detection head, MD3D requires few hand-crafted eliminates inefficiency separating regression channel class, thus offering both memory benefits. is designed utilize types feature encoding; therfore, it applied flexibly replacing head existing Experimental results KITTI Waymo open datasets show proposed method outperforms its counterparts based conventional overall performance, latency, memory. code publicly available at https://github.com/sky77764/MD3D.