# LSK3DNet: Towards Effective and Efficient 3D Perception with Large Sparse Kernels

**Authors:** Tuo Feng, Wenguan Wang, Fan Ma, Yi Yang  
**arXiv:** [2403.15173](https://arxiv.org/abs/2403.15173)  
**Date:** 2024-03-22  
**Summary page:** https://www.summarizepaper.com/en/paper/2403.15173v1/

## Key points

- LSK3DNet is a novel approach for autonomous systems to process large-scale point cloud data efficiently
- It addresses challenges of processing sparse and irregular point clouds with limited compute resources
- Core innovation lies in dynamic pruning techniques for amplifying 3D kernel size
- Components include Spatial-wise Dynamic Sparsity (SDS) and Channel-wise Weight Selection (CWS)
- Outperforms classical models and large kernel designs on benchmark datasets
- Achieves state-of-the-art performance on SemanticKITTI dataset
- Reduces model size by 40% and computing operations by 60% compared to naive large 3D kernel models

## Summary

, , , , 


LSK3DNet: An Efficient Solution for Autonomous Systems to Process Large-Scale Point Cloud Data with Limited Compute Resources

 is a novel approach that addresses the challenges faced by autonomous systems in processing large-scale, sparse, and irregular point clouds with limited compute resources. Its key objective is to develop efficient and effective LiDAR perception methods capable of handling these complex datasets. To improve performance, traditional methods often increase the 3D kernel size, resulting in a significant increase in computational overhead. However, LSK3DNet introduces a streamlined 3D large kernel design that eliminates redundant weights and effectively utilizes larger kernels. The core innovation of LSK3DNet lies in its utilization of dynamic pruning techniques to amplify the 3D kernel size. This method consists of two main components: Spatial-wise Dynamic Sparsity (SDS) and Channel-wise Weight Selection (CWS). SDS dynamically prunes and regrows volumetric weights from the outset to learn a large sparse 3D kernel, enhancing performance while reducing model size and computational cost significantly. Additionally, CWS selects crucial channels for 3D convolution during training and subsequently prunes redundant channels to accelerate inference for 3D vision tasks. Experiments on three benchmark datasets and five tracks demonstrate the effectiveness of LSK3DNet, outperforming classical models and large kernel designs. Notably, it achieves state-of-the-art performance on SemanticKITTI with accuracy rates of 75.6% on single-scan and 63.4% on multi-scan tasks. Furthermore, compared to naive large 3D kernel models, LSK3DNet reduces model size by 40% and computing operations by 60%. In summary,  represents a significant advancement in LiDAR perception methods, offering an efficient and effective solution for processing large-scale point cloud data with limited compute resources. Its innovative use of dynamic pruning techniques sets it apart from traditional approaches, making it a promising tool for enhancing autonomous systems' capabilities in handling complex environmental data efficiently.