Research map / Point clouds / Point cloud classification
DiffLO: Semantic-Aware LiDAR Odometry with Diffusion-Based Refinement
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 DiffLO, a LiDAR odometry network that distills semantic knowledge from a frozen 3D segmentation teacher and refines poses with a conditional diffusion model, achieving state-of-the-art learning-based results on KITTI.
Problem
LiDAR odometry estimates relative pose between consecutive point clouds, but learning-based methods struggle with dynamic objects, repetitive structures, low texture, and noise, and their deterministic regression gives no uncertainty estimate.
Approach
Combines knowledge distillation of semantic features from a pretrained 3D segmentation model into the odometry student (used only in training) with a diffusion-based iterative residual pose refinement, conditioned on geometry, cross-frame cost volume, semantic, and coarse pose embeddings.
Key results
- On KITTI 07-10 average, DiffLO reports t_rel 0.708% and r_rel 0.328 deg/100m, versus EfficientLO 0.795% and 0.395.
- Removing the diffusion refinement raises mean t_rel on 07-10 from 0.708% to 5.995%.
- Inference takes 76.9 ms per frame, under the 100 ms real-time threshold.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the Semantic Perception Module (SPM) (ablation) | DiffLO · KITTI 07-10 mean | trel (%): 0.793 -> 0.708 | Helped |
| add the Semantic Perception Module (SPM) (ablation) | DiffLO · KITTI 07-10 mean | rrel (°/100m): 0.398 -> 0.328 | Helped |
| add diffusion-based pose refinement (ablation) | DiffLO · KITTI 07-10 mean | trel (%): 5.995 -> 0.708 | Helped |
| add diffusion-based pose refinement (ablation) | DiffLO · KITTI 07-10 mean | rrel (°/100m): 3.265 -> 0.328 | Helped |
| use GRU-based Mθ instead of trans-based Mθ (ablation) | DiffLO · KITTI 07-10 mean | trel (%): 0.888 -> 0.708 | Helped |
| use GRU-based Mθ instead of ptrans-based Mθ (ablation) | DiffLO · KITTI 07-10 mean | trel (%): 0.725 -> 0.708 | Helped |
| add the cost volume condition (ablation) | DiffLO · KITTI 07-10 mean | trel (%): 0.860 -> 0.708 | Helped |
| add the cost volume condition (ablation) | DiffLO · KITTI 07-10 mean | rrel (°/100m): 0.518 -> 0.328 | Helped |
| add the geometry feature condition (ablation) | DiffLO · KITTI 07-10 mean | trel (%): 0.787 -> 0.708 | Helped |
| add the semantic feature condition (ablation) | DiffLO · KITTI 07-10 mean | trel (%): 0.798 -> 0.708 | Helped |
| add the semantic feature condition (ablation) | DiffLO · KITTI 07-10 mean | rrel (°/100m): 0.395 -> 0.328 | Helped |
| add the coarse pose embedding condition (ablation) | DiffLO · KITTI 07-10 mean | trel (%): 0.750 -> 0.708 | Helped |
Limitations
- Evaluated only on KITTI odometry, with a single dataset and sequence split.
- Trained on sequences 00-06 only, and the paper notes that the training/testing splits differ across compared methods.
- Compared against A-LOAM without mapping, but not against full mapping-based SLAM on all sequences.
Experimental setup
Datasets: KITTI odometry
Benchmarks: KITTI odometry
Metrics: translational RMSE (%) rotational RMSE (deg/100m)
Baselines: ICP-po2po ICP-po2pl GICP CLS LOAM A-LOAM LO-Net DELO PWCLO-Net TransLO EfficientLO LodoNet
Builds on: PWCLO-Net PointConv DDPM Point Transformer-based cost volume refinement A-LOAM
Closest papers by meaning
- DiffLoc: Diffusion Model for Outdoor LiDAR Localization (CVPR 2024 · similarity 0.89)
- DiffuBox: Refining 3D Object Detection with Point Diffusion (NeurIPS 2024 · similarity 0.87)
- LiSA: LiDAR Localization with Semantic Awareness (CVPR 2024 · similarity 0.85)
- OLiDM: Object-aware LiDAR Diffusion Models for Autonomous Driving (AAAI 2025 · similarity 0.85)
- Is Your LiDAR Placement Optimized for 3D Scene Understanding? (NeurIPS 2024 · similarity 0.85)
- TALoS: Enhancing Semantic Scene Completion via Test-time Adaptation on the Line of Sight (NeurIPS 2024 · similarity 0.84)
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.