This paper proposes a 3D Structural Convolutional Network (3D-SCN) for 3D convolutional encoding layers in LiDAR-based self-driving applications. The 3D-SCN leverages novel convolutional kernels that incorporate cosine similarity and Euclidean distance metrics to adeptly capture geometric characteristics from LiDAR datasets. This design is specifically crafted to maintain feature invariance amidst the disparities in regional data and sensor-specific channel variations. Experiment conducted on various LiDAR-based point cloud datasets demonstrate that the proposed 3D-SCN (3D Structural Convolutional Network) shows consistent performance across different LiDAR sensor specifications, even when trained on a specific dataset. To further validate its effectiveness and enhance the diversity of the LiDAR domain, we introduce the PanKyo dataset, which includes a comprehensive set of samples with 32, 64, and 128 channel domain differences. The results presented underscore the efficacy of the 3D-SCN in enhancing performance and robustness for LiDAR-based 3D recognition tasks in the context of self-driving applications.
This work was supported by a Korea Agency for Infrastructure Technology Advancement (KAIA) grant funded by the Korean government (MOLIT) (RS-2022-001560), the BK21 FOUR program of the National Research Foundation Korea funded by the Ministry of Education (NRF5199991014091), and partially supported by the Convergence and Open Sharing System (Future Automotive Sector) by the Ministry of Education and the National Research Foundation of Korea.
