文献类型：期刊文献

英文题名：Keypoints Representation of Density-aware and the Spatial-Channel-wise Decoder for 3D Object Detection

作者：Ma, Qiming[1]; Zhu, Yu[1]

机构：[1] School of Information Science and Engineering, East China University of Science and Technology, Shanghai, China

年份：2022

外文期刊名：Proceedings - 2022 15th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics, CISP-BMEI 2022

收录：EI(收录号：20230313390464)

语种：英文

外文关键词：3D modeling - Channel coding - Decoding - Object recognition - Statistical tests

摘要：Current 3D object detection frameworks based on LiDAR mainly used sparse convolution as the backbone network after voxelization to extract features, and applied grids to further refine the proposal boxes. While these operations may limit the accuracy improvement of 3D object detection, because they destroyed the geometric characteristics of point clouds to a large extent, including density and object shape. Therefore, in this paper, we proposed a method to refine the proposals by estimating density-aware information in the second stage. A certain number of key points were sampled in each proposal, and then applied the self-attention module to study the relations between these key points. Then the designed spatial-channel-wise decoder fused channel-wise and spatial-wise features to obtain the global representation of the object. Finally, the global representation was fed into the detect head to obtain a more accurate box. The performance of our proposed 3D detection model was evaluated on the KITTI dataset, and the average accuracy of car class on the test set and validation split was 80.62% and 85.39% respectively, and the average accuracy of three classes in KITTI on the validation split was 72.41%. ? 2022 IEEE.