Homogenous multimodal 3D object detection based on deformable Transformer and attribute dependencies

In this work, we propose a unified multi-representation 3D object detection framework called attribute dependent Homogenous fusion 3D object Detection Network ADHF. The proposed method aims to solve the problem of inefficient utilization of point cloud multiple representations. For this purpose, the cross-head self-attention mechanism CSA extracts key points and Voxels features, and it realizes cross-head information interaction using only multi-head attention and channel adjustment. Different from previous work, our method mines the similarity between Voxels features and key features in Deformable Attention Transformer by improving the data transmission information flow, so as to capture remote dependencies, aggregate multi-representation information, and make extracted point voxels features more discriminating. The previous methods usually predict the properties of 3D objects at one time, and fail to make full use of the hidden sequence information in the prediction features of objects. The ADHF framework is equipped with a novel attribute dependent detection head ADH to solve the above problems. Overall, ADHF is an interim attempt to represent different modes in a unified framework. The proposed method achieves leading performance for target detection in Kitti set.