Applying Bilateral Guided Multi-Viewed Fusion on Asymmetrical 3D Convolution Networks for 3D LiDAR semantic segmentation
Tai Huu - Phuong Tran, Jae Wook Jeon · 2022 IEEE International Conference on Consumer Electronics-Asia (ICCE-Asia) · 2022
We present a novel approach for 3D semantic segmentation using LiDAR sensor. In this work, we focus on improving the accuracy of the cylindrical partition and asymmetrical 3D convolution networks with a modification on their dimension-decomposition based context modeling module and asymmetrical residual block. The initial version simply performs multiplication and addition operators to combine two feature branches. In our modification, we apply the bilateral guided multi-viewed fusion module to provide better features for the classification stages. We trained and tested our model on SemanticKITTI dataset, our implementation improves better accuracy than the default asymmetrical 3D convolution networks, in which uses cylindrical voxel for the point cloud representation.