Fig. 2
From: CenterTransFuser: radar point cloud and visual information fusion for 3D object detection
CenterTransFuser model consists of two branches, namely radar and image. The image branch, after feature extraction by the backbone network, generates preliminary 3D information and 2D regions of interest for the radar branch. The radar branch is supplemented by data processing and frustum association with the information provided by the image. The information from both branches is fed into the cross-transformer module for contextual interaction and into the detection head for classification and regression