Cylindrical and Asymmetrical 3D Convolution Networks for LiDAR-Based Perception.

Area of Science:

• Computer Vision
• Robotics
• Autonomous Driving

Background:

• Current LiDAR perception methods often project 3D point clouds to 2D, losing crucial 3D geometric information.
• While 3D voxelization and convolution exist, they show limited improvement on sparse, outdoor LiDAR data.

Purpose of the Study:

• To develop a new framework for outdoor LiDAR segmentation that effectively utilizes 3D geometric patterns.
• To address the challenges of sparsity and varying density in outdoor point clouds.
• To create a versatile backbone for downstream tasks like semantic segmentation, panoptic segmentation, and 3D detection.

Main Methods:

• Proposed a framework featuring cylindrical partition and asymmetrical 3D convolution networks.
• Designed to explore 3D geometric patterns while preserving inherent properties of outdoor point clouds.
• Evaluated the model as a backbone for semantic segmentation, panoptic segmentation, and 3D detection.

Main Results:

• Achieved state-of-the-art results on the SemanticKITTI leaderboard for semantic segmentation (single-scan and multi-scan).
• Significantly outperformed existing methods on nuScenes and A2D2 datasets for semantic segmentation.
• Demonstrated strong performance and good generalization for LiDAR panoptic segmentation and 3D detection.

Conclusions:

• The proposed 3D framework effectively handles outdoor LiDAR data challenges.
• The method offers a competitive and generalizable solution for various LiDAR-based perception tasks.
• This work advances the state-of-the-art in autonomous driving perception systems.