Research map / Point clouds / Segmentation and detection
LiSA: LiDAR Localization with Semantic Awareness
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.
Adds semantic awareness to scene coordinate regression LiDAR localization by diffusion-based distillation from a frozen 3D segmentation model, with no inference overhead.
Problem
LiDAR localization with scene coordinate regression treats all points equally, though dynamic objects and repetitive structures hurt localization; segmentation-based filtering adds inference overhead and is sensitive to noisy labels.
Approach
Distill per-point semantic features from a frozen 3D segmentation teacher into an SCR student via a semantic regression branch and diffusion-based distillation loss (DDPM denoising plus L1), then discard teacher and distillation modules at inference.
Key results
- On QEOxford, LiSA reaches 0.95m / 1.14° average error, a 38% position and 29% orientation improvement over SGLoc.
- On NCLT, LiSA achieves 1.51m / 2.34° average, improving on SGLoc by 17% position and 34% orientation.
- LiSA runs at 38ms on Oxford and 75ms on NCLT with no extra inference cost over the SCR baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add semantic awareness through knowledge distillation | SGLoc · QEOxford | position error: 1.53m -> 0.95m | Helped |
| add semantic awareness through knowledge distillation | SGLoc · QEOxford | orientation error: 1.60◦ -> 1.14◦ | Helped |
| add semantic awareness through knowledge distillation | SGLoc · Oxford | position error: 3.14m -> 2.74m | Helped |
| add semantic awareness through knowledge distillation | SGLoc · Oxford | orientation error: 1.88◦ -> 1.32◦ | Hurt |
| add semantic awareness through knowledge distillation | SGLoc · NCLT | position error: 1.83m -> 1.51m | Helped |
| add semantic awareness through knowledge distillation | SGLoc · NCLT | orientation error: 3.54◦ -> 2.34◦ | Helped |
| remove semantic information (ablation) | LiSA on QEOxford · QEOxford | position error: 0.95m -> 1.79m | Hurt |
| replace knowledge distillation with point filtering (ablation) | LiSA on QEOxford · QEOxford | position error: 0.95m -> 1.20m | Hurt |
| replace knowledge distillation with point filtering (ablation) | LiSA on QEOxford · QEOxford | orientation error: 1.14◦ -> 1.97◦ | Helped |
| replace knowledge distillation with feature concatenation (ablation) | LiSA on QEOxford · QEOxford | position error: 0.95m -> 1.14m | Hurt |
| replace knowledge distillation with feature concatenation (ablation) | LiSA on QEOxford · QEOxford | orientation error: 1.14◦ -> 1.60◦ | Hurt |
| use DDPM instead of L1 distillation (ablation) | LiSA with SPVNAS features · QEOxford | position error: 1.28m -> 1.15m | Helped |
Surprising findings
- Naive point filtering by predicted semantic labels can improve position error while worsening orientation error, showing noisy labels hurt consistency.
Limitations
- Segmentation teacher performs much worse on NCLT, limiting semantic quality there.
- Orientation accuracy on the non-enhanced Oxford dataset still trails the best APR method HypLiLoc.
Experimental setup
Datasets: Oxford Radar RobotCar QEOxford NCLT
Benchmarks: QEOxford Oxford Radar RobotCar NCLT
Metrics: mean position error (m) mean orientation error (deg) inference time (ms)
Baselines: SGLoc PNVLAD DCP PointLoc PosePN++ STCLoc HypLiLoc SGLoc+PGO
Builds on: SGLoc DDPM knowledge distillation SphereFormer
Closest papers by meaning
- LightLoc: Learning Outdoor LiDAR Localization at Light Speed (CVPR 2025 · similarity 0.87)
- L2RSI: Cross-view LiDAR-based Place Recognition for Large-scale Urban Scenes via Remote Sensing Imagery (NeurIPS 2025 · similarity 0.86)
- R-SCoRe: Revisiting Scene Coordinate Regression for Robust Large-Scale Visual Localization (CVPR 2025 · similarity 0.86)
- DiffLoc: Diffusion Model for Outdoor LiDAR Localization (CVPR 2024 · similarity 0.85)
- DiffLO: Semantic-Aware LiDAR Odometry with Diffusion-Based Refinement (CVPR 2025 · similarity 0.85)
- DeepPointMap: Advancing LiDAR SLAM with Unified Neural Descriptors (AAAI 2024 · similarity 0.85)
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.