Research map / Point clouds / Segmentation and detection
TerraSeg: Self-Supervised Ground Segmentation for Any LiDAR
Ted Lentsch, Santiago Montiel-Marín, Holger Caesar, Dariu M. Gavrila
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add pre-processing (0.5% low-height quantile filter) to PseudoLabeler (ablation) | PseudoLabeler · nuScenes val | mIoU: 79.40 -> 86.38 | Improved |
| add pre-processing to PseudoLabeler (ablation) | PseudoLabeler · Waymo Perception val | mIoU: 85.00 -> 85.47 | Improved |
| add pillar-based post-processing to PseudoLabeler (ablation) | PseudoLabeler · SemanticKITTI val | mIoU: 83.82 -> 91.97 | Improved |
| add pillar-based post-processing to PseudoLabeler (ablation) | PseudoLabeler · nuScenes val | mIoU: 86.38 -> 90.63 | Improved |
| increase recovery threshold tau from 0.05 to 0.20 (ablation) | PseudoLabeler post-processing · avg mIoU, pillar 0.50 m | mIoU: 91.85 -> 89.76 | Hurt |
| change pillar size from 0.50 m to 0.25 m or 1.00 m (ablation) | PseudoLabeler post-processing · avg mIoU, tau 0.05 | mIoU: 91.85 -> 91.08 / 90.63 | Hurt |
| train on pseudo-labels from many LiDARs instead of single-dataset supervised GndNet | TerraSeg-B vs GndNet · Waymo Perception val | mIoU: 61.83 -> 94.40 | Improved |
| use manual labels (supervised oracle) instead of pseudo-labels | TerraSeg-B · nuScenes val | mIoU: 92.47 -> 95.31 | Improved |
| use larger TerraSeg-B instead of TerraSeg-S | TerraSeg · nuScenes val | mIoU: 91.45 -> 92.47 | Improved |
| distill PseudoLabeler into TerraSeg-B model | PseudoLabeler · nuScenes val | mIoU: 90.63 -> 92.47 | Improved |
| non-flat terrain (sigma > 0.40 m) vs flat | TerraSeg-B · nuScenes val | mIoU: 93.12 -> 87.81 | Hurt |
Closest papers by meaning
- Walking Your LiDOG: A Journey Through Multiple Domains for LiDAR Semantic Segmentation (ICCV 2023 · similarity 0.90)
- MemorySeg: Online LiDAR Semantic Segmentation with a Latent Memory (ICCV 2023 · similarity 0.89)
- Better Call SAL: Towards Learning to Segment Anything in Lidar (ECCV 2024 · similarity 0.89)
- Learnable Earth Parser: Discovering 3D Prototypes in Aerial Scans (CVPR 2024 · similarity 0.88)
- TREND: Unsupervised 3D Representation Learning via Temporal Forecasting for LiDAR Perception (NeurIPS 2025 · similarity 0.87)
- TASeg: Temporal Aggregation Network for LiDAR Semantic 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.
