Research map / Point clouds / Point cloud classification
WeatherGen: A Unified Diverse Weather Generator for LiDAR Point Clouds via Spider Mamba Diffusion
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 WeatherGen, a unified diffusion framework that generates diverse-weather (snow, fog, rain) LiDAR range maps using a spider mamba denoiser, real-data latent alignment, and CLIP-supervised contrastive weather control, and builds a mini-weather dataset for 3D detection.
Problem
Collecting adverse-weather LiDAR data is costly, and existing LiDAR simulators handle only one weather type with a single non-learnable physical model and limited fidelity.
Approach
A diffusion model over range maps whose denoiser (spider mamba) scans point-level along LiDAR beam circles and central rays rather than image patches, combined with a learnable map-based data producer for pre-training, a KL-based latent feature aligner using real data, and a contrastive CLIP-language weather controller.
Key results
- On KITTI-360 unconditional generation, WeatherGen reaches FRD 138.62 vs 149.66 for R2DM and 164.16 for Text2LiDAR, and the best BEV MMD (0.39e-4).
- On Seeing Through Fog weather-conditional generation, WeatherGen (Snow) achieves FPD 59.28 and FRD 124.17, outperforming LSS (Snow) at FPD 106.37.
- Training PointPillars on mini-weather data (7.4% of training frames) improves dense-fog Car AP@.5IoU from 35.98 to 42.14 over no augmentation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use WeatherGen instead of Text2LiDAR | unconditional LiDAR generation · KITTI-360 | FRD: 164.16 -> 138.62 | Helped |
| use WeatherGen instead of Text2LiDAR | unconditional LiDAR generation · KITTI-360 | FPD: 4.81 -> 6.15 | Hurt |
| use WeatherGen instead of the weather-specific simulator | weather-conditional LiDAR generation · Seeing Through Fog snow | FPD: 106.37 -> 59.28 | Helped |
| use WeatherGen instead of the weather-specific simulator | weather-conditional LiDAR generation · Seeing Through Fog fog | FPD: 319.32 -> 314.14 | Helped |
| use WeatherGen instead of the weather-specific simulator | weather-conditional LiDAR generation · Seeing Through Fog rain | FPD: 301.11 -> 86.40 | Helped |
| add learnable MDP (ablation) | WeatherGen · Seeing Through Fog snow heavy test split | FPD: 84.58 -> 59.28 | Helped |
| use spider mamba generator instead of U-Net (ablation) | WeatherGen · Seeing Through Fog snow heavy test split | FPD: 79.45 -> 59.28 | Helped |
| add contrastive learning-based controller (ablation) | WeatherGen · Seeing Through Fog snow heavy test split | FPD: 66.35 -> 59.28 | Helped |
| add latent feature aligner (ablation) | WeatherGen · Seeing Through Fog snow heavy test split | FPD: 77.39 -> 59.28 | Helped |
| fine-tune using real-world data (ablation) | WeatherGen · Seeing Through Fog snow heavy test split | FPD: 148.25 -> 59.28 | Helped |
Surprising findings
- Fog generation (FPD 314.14) is not better than the simulator LSS-style baselines on point cloud metric, indicating weakness on fog data relative to snow and rain.
Limitations
- Fog generation performance is weaker than other weather types on point-cloud metrics.
- Adverse-weather training data is scarce, so fine-tuning relies on small Seeing Through Fog splits.
- Only snow, fog, and rain are covered.
Experimental setup
Datasets: KITTI-360 Seeing Through Fog
Benchmarks: KITTI-360 Seeing Through Fog
Metrics: FPD FRD MMD JSD Car AP Cyclist AP Pedestrian AP
Baselines: LiDARGen LiDM R2DM Text2LiDAR LiDARGAN LiDARVAE ProjectedGAN LSS FSRL LISA DROR
Builds on: DDPM Mamba Vision Mamba UltraLight VM-UNet CLIP LiDARGen
Closest papers by meaning
- Towards foundational LiDAR world models with efficient latent flow matching (NeurIPS 2025 · similarity 0.87)
- UniMix: Towards Domain Adaptive and Generalizable LiDAR Semantic Segmentation in Adverse Weather (CVPR 2024 · similarity 0.86)
- Sunshine to Rainstorm: Cross-Weather Knowledge Distillation for Robust 3D Object Detection (AAAI 2024 · similarity 0.86)
- Towards Realistic Scene Generation with LiDAR Diffusion Models (CVPR 2024 · similarity 0.86)
- An Empirical Study of the Generalization Ability of Lidar 3D Object Detectors to Unseen Domains (CVPR 2024 · similarity 0.85)
- TREND: Unsupervised 3D Representation Learning via Temporal Forecasting for LiDAR Perception (NeurIPS 2025 · 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.