Research map / Point clouds / Segmentation and detection
HiLoTs: High-Low Temporal Sensitive Representation Learning for Semi-Supervised LiDAR Segmentation in Autonomous Driving
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 HiLoTs, a semi-supervised LiDAR segmentation method that models high- and low-temporal-sensitivity flows over continuous frames, fused via cross-attention and trained with a mean-teacher framework.
Problem
Semi-supervised LiDAR semantic segmentation for autonomous driving typically uses spatial cues or only short-term (two-frame) temporal information, missing long-term temporal dynamics that differ between near and distant objects.
Approach
Splits cylindrical voxel features by distance into a high temporal sensitivity flow (distant voxels, self-attention) and a low temporal sensitivity flow (nearby voxels, self-attention plus cross-attention with the high flow), aggregates neighboring voxels into super-voxels to cut attention cost, and embeds the result into a Mean Teacher framework.
Key results
- On SemanticKITTI with 1% labels, HiLoTs reaches 58.6 mIoU versus 45.4 for fully supervised Cylinder3D (reproduced).
- On SemanticKITTI at 10% labels, HiLoTs achieves 65.7 mIoU, outperforming LiDAR-only DDSemi (65.1) and FRNet (64.8).
- On nuScenes at 50% labels, HiLoTs achieves 76.9 mIoU, close to LiDAR+camera LaserMix++ (76.3) and above CyMix+IPSL (80.5 is higher, so comparable rather than superior).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add HTSF and LTSF rather than omit HEU (ablation) | HiLoTs · SemanticKITTI val, 10% labels | mIoU: 59.2 -> 65.7 | Helped |
| add LTSF to HTSF (ablation) | HiLoTs · SemanticKITTI val, 10% labels | mIoU: 63.4 -> 65.7 | Helped |
| add HTSF to LTSF (ablation) | HiLoTs · SemanticKITTI val, 10% labels | mIoU: 62.8 -> 65.7 | Helped |
| use cross-attention with LTSF as query rather than addition (ablation) | HiLoTs HEU · SemanticKITTI val, 10% labels | mIoU: 63.6 -> 65.7 | Helped |
| use cross-attention with LTSF as query rather than concatenation (ablation) | HiLoTs HEU · SemanticKITTI val, 10% labels | mIoU: 64.3 -> 65.7 | Helped |
| use LTSF as query instead of HTSF (ablation) | HiLoTs HEU · SemanticKITTI val, 10% labels | mIoU: 65.4 -> 65.7 | Helped |
| use multi-voxel aggregation instead of random voxel selection (ablation) | HiLoTs MVA · SemanticKITTI val, 10% labels | mIoU: 63.5 -> 65.7 | Helped |
| use multi-voxel aggregation instead of density-based voxel selection (ablation) | HiLoTs MVA · SemanticKITTI val, 10% labels | mIoU: 63.9 -> 65.7 | Helped |
| use cylindrical rather than cubic voxelization (ablation) | HiLoTs backbone · SemanticKITTI val, 10% labels | mIoU: 64.3 -> 65.7 | Helped |
| use cylindrical rather than pillar voxelization (ablation) | HiLoTs backbone · SemanticKITTI val, 10% labels | mIoU: 62.4 -> 65.7 | Helped |
| use cylindrical rather than spherical voxelization (ablation) | HiLoTs backbone · SemanticKITTI val, 10% labels | mIoU: 64.8 -> 65.7 | Helped |
| set EMA ratio to 0.99 rather than 0.5 (ablation) | HiLoTs · SemanticKITTI val, 10% labels | mIoU: 65.2 -> 65.7 | Helped |
Surprising findings
- Changing the EMA ratio between 0.5 and 0.99 has only modest impact, unlike prior mean-teacher works that report high sensitivity.
- Using HTSF (distant-region flow) alone gives larger gains than LTSF alone, despite LTSF targeting the stable near region.
Limitations
- Designed specifically for autonomous driving sequences; not suitable for general object point cloud segmentation such as PointNet or KPConv settings.
- Requires multi-frame (t=5) input, increasing memory and input requirements.
- Weaker performance on classes such as other-ground, bicycle, and traffic-sign.
Experimental setup
Datasets: SemanticKITTI nuScenes SemanticKITTI-C nuScenes-C
Benchmarks: SemanticKITTI nuScenes SemanticKITTI-C nuScenes-C
Metrics: mIoU IoU per class
Baselines: Cylinder3D RangeViT SphereFormer MarS3D GPC PolarMix LaserMix LiM3D ImageTo360 IGNet DDSemi FRNet
Builds on: Mean Teacher Cylinder3D Minkowski-UNet LaserMix
Closest papers by meaning
- Temporal Consistent 3D LiDAR Representation Learning for Semantic Perception in Autonomous Driving (CVPR 2023 · similarity 0.91)
- VLM2Scene: Self-Supervised Image-Text-LiDAR Learning with Foundation Models for Autonomous Driving Scene Understanding (AAAI 2024 · similarity 0.88)
- TREND: Unsupervised 3D Representation Learning via Temporal Forecasting for LiDAR Perception (NeurIPS 2025 · similarity 0.88)
- ICP-Flow: LiDAR Scene Flow Estimation with ICP (CVPR 2024 · similarity 0.87)
- Exploring Scene Affinity for Semi-Supervised LiDAR Semantic Segmentation (CVPR 2025 · similarity 0.87)
- Multi-Space Alignments Towards Universal LiDAR Segmentation (CVPR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.