Universal LiDAR segmentation

Paper Title:

Multi-Space Alignments Towards Universal LiDAR Segmentation

Published on:

2 May 2024

Primary Category:

Computer Vision and Pattern Recognition

Paper Authors:

Youquan Liu,

Lingdong Kong,

Xiaoyang Wu,

Runnan Chen,

Xin Li,

Liang Pan,

Ziwei Liu,

Yuexin Ma

Bullets

Key Details

•

Combines multi-task, multi-dataset, and multi-modality data

•

Conducts alignments in data, feature, and label spaces

•

Achieves state-of-the-art segmentation performance

•

Uses a single set of parameters across diverse datasets

•

Demonstrates strong robustness and generalizability

Explore the topics in this paper

driving datasets

lidar segmentation

multi-dataset learning

multi-modality learning

multi-task learning

AI generated summary

Universal LiDAR segmentation

This paper presents M3Net, a framework for multi-task, multi-dataset, multi-modality LiDAR segmentation using a single set of parameters. It combines large-scale driving datasets with different sensors and conducts alignments in data, feature, and label spaces during training. This allows M3Net to train state-of-the-art segmentation models by taming heterogeneous data. Experiments on 12 datasets show it achieves top performance on SemanticKITTI, nuScenes, and Waymo using one shared parameter set.