Research map / Point clouds / Segmentation and detection

LiMoE: Mixture of LiDAR Representation Learners from Automotive Scenes

CVPR 2025 · Point clouds › Segmentation and detection · approach: LiDAR segmentation for driving · 9 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes LiMoE, which uses a mixture-of-experts layer over range-view, sparse-voxel, and point LiDAR representations to distill combined features into a single 3D network during pretraining and to fuse semantic logits downstream.

Problem

Existing LiDAR image-to-LiDAR pretraining relies mainly on sparse voxel representations and ignores the complementary attributes of range-view and raw-point representations.

Approach

Integrates range images, sparse voxels, and raw points via a per-point gated Mixture-of-Experts layer, first in contrastive pretraining (CML, distilling mixed features into one network) and then in downstream semantic logit fusion (SMS).

Key results

• With SuperFlow and ViT-L image backbone, LiMoE reaches 49.35% mIoU on nuScenes 1% linear probing versus 48.01% for SuperFlow alone (and up to 77.85% full fine-tuning).
• On Robo3D nuScenes-C, LiMoE with SuperFlow reduces mCE from 91.67 to 88.43 and raises mRR from 83.17 to 83.28 under full fine-tuning.
• Ablation with three identical voxel experts gives only 42.72 mIoU on nuScenes 1% vs 46.89 for heterogeneous LiMoE experts, indicating representation diversity matters.

Measured effects reported in the paper

Surprising findings

• Using three identical sparse-voxel experts yields only a minor gain over single-representation pretraining, showing that the gains come mainly from representation diversity.

Limitations

• Downstream SMS adds computational cost from running three representation backbones.
• Range-view experts cannot be used on most datasets because sensor FoV parameters are unknown, so only voxel and point experts are used there.

Experimental setup

Datasets: nuScenes SemanticKITTI Waymo Open ScribbleKITTI RELLIS-3D SemanticPOSS SemanticSTF SynLiDAR DAPS-3D Synth4D Robo3D (nuScenes-C)

Benchmarks: nuScenes SemanticKITTI Waymo Open ScribbleKITTI RELLIS-3D SemanticPOSS SemanticSTF SynLiDAR DAPS-3D Synth4D nuScenes-C

Metrics: mIoU mCE mRR

Baselines: SLidR Seal SuperFlow PPKT TriCC CSC HVDistill Random initialization

Builds on: SLidR SuperFlow Seal Mixture of Experts

Closest papers by meaning

• TREND: Unsupervised 3D Representation Learning via Temporal Forecasting for LiDAR Perception (NeurIPS 2025 · similarity 0.87)
• VLM2Scene: Self-Supervised Image-Text-LiDAR Learning with Foundation Models for Autonomous Driving Scene Understanding (AAAI 2024 · similarity 0.87)
• Exploring Scene Affinity for Semi-Supervised LiDAR Semantic Segmentation (CVPR 2025 · similarity 0.86)
• PC-BEV: An Efficient Polar-Cartesian BEV Fusion Framework for LiDAR Semantic Segmentation (AAAI 2025 · similarity 0.86)
• HiLoTs: High-Low Temporal Sensitive Representation Learning for Semi-Supervised LiDAR Segmentation in Autonomous Driving (CVPR 2025 · similarity 0.86)
• Multi-Space Alignments Towards Universal LiDAR Segmentation (CVPR 2024 · similarity 0.86)