Semantic-enhanced Graph Voxelization for Pillar-based 3D Detection from Point Clouds
Yonglin Tian, Xiao Wang, Yu Shen, Yuhang Liu, Zilei Wang, Fei–Yue Wang · 2022 IEEE 25th International Conference on Intelligent Transportation Systems (ITSC) · 2022
Pillar-based 3D detection simplifies the processing pipelines of feature extraction for irregular point clouds. However, the rigid voxelization adopted by pillar-based methods, inevitably and irreversibly, cuts off the direct connection between plenty of point pairs, thus hindering the modeling of local structures. Besides, it is difficult to handle intractable situations in 3D detection using structural information alone. In this paper, we propose a semantic-enhanced graph voxelization method to improve pillar-based feature extraction from two aspects, i.e., the construction of pillars and the learning of pillar-level features. For pillar construction, a novel voxelization method is proposed to capture more sufficient local structures with graph-based modeling of the relationships between adjacent point pairs. To learn representative features for each pillar, semantic enhancement is adopted by learning context awareness and class awareness pixel by pixel with a front-view-based segmentation branch. Extensive experiments conducted on the KITTI dataset demonstrate the effectiveness of the proposed method. Compared with the baseline, our method achieves 2.81 points and 2.65 points higher performance in 3D detection on KITTI test set for cars and cyclists individually.