Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution
Name
2007.16100.pdf
Description
Accepted version
Size
4.74 MB
Format
Adobe PDF
Checksum (MD5)
4f5ecc448b315435fd2da64f1eda6ee7
Author(s)
Tang, Haotian
Liu, Zhijian
Zhao, Shengyu
Lin, Yujun
Lin, Ji
Wang, Hanrui
Han, Song
Date Issued
2020
Journal
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Publisher
Springer International Publishing
Citation
Tang, Haotian, Liu, Zhijian, Zhao, Shengyu, Lin, Yujun, Lin, Ji et al. 2020. "Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution." Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 12373.
Version
Author's final manuscript
Abstract
© 2020, Springer Nature Switzerland AG. Self-driving cars need to understand 3D scenes efficiently and accurately in order to drive safely. Given the limited hardware resources, existing 3D perception models are not able to recognize small instances (e.g., pedestrians, cyclists) very well due to the low-resolution voxelization and aggressive downsampling. To this end, we propose Sparse Point-Voxel Convolution (SPVConv), a lightweight 3D module that equips the vanilla Sparse Convolution with the high-resolution point-based branch. With negligible overhead, this point-based branch is able to preserve the fine details even from large outdoor scenes. To explore the spectrum of efficient 3D models, we first define a flexible architecture design space based on SPVConv, and we then present 3D Neural Architecture Search (3D-NAS) to search the optimal network architecture over this diverse design space efficiently and effectively. Experimental results validate that the resulting SPVNAS model is fast and accurate: it outperforms the state-of-the-art MinkowskiNet by 3.3%, ranking 1 on the competitive SemanticKITTI leaderboard. It also achieves 8–23 computation reduction and 3 measured speedup over MinkowskiNet and KPConv with higher accuracy. Finally, we transfer our method to 3D object detection, and it achieves consistent improvements over the one-stage detection baseline on KITTI.
MIT Department
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
Terms of Use
Creative Commons Attribution-Noncommercial-Share Alike
Persistent DSpace Link
DOI of Published Version
https://doi.org/10.1007/978-3-030-58604-1_41
