Research map / 3D vision / 3D generation
Spiral: Semantic-Aware Progressive LiDAR Scene Generation and Understanding
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.
Spiral jointly generates depth, reflectance, and semantic label range-view LiDAR scenes with a single diffusion model using progressive EMA semantic prediction and closed-loop conditional/unconditional inference, plus semantic-aware generation metrics.
Problem
Range-view LiDAR diffusion models generate only depth and reflectance without semantic labels; two-step pipelines with post-hoc segmentation give poor cross-modal consistency, and existing evaluation metrics ignore semantics.
Approach
A single range-view diffusion model (Efficient U-Net backbone) with a semantic head that predicts label maps during unconditional steps, smooths them with EMA across denoising steps, and feeds confident semantic maps back as conditions in a closed-loop alternation; semantic-aware FRD/FPD/MMD/JSD metrics extend evaluation to labeled scenes.
Key results
- On SemanticKITTI, Spiral (61M params, 256 NFE) outperforms R2DM+SPVCNN++ by 31.03% on S-FRD, 56.33% on S-FPD, and 50.94% on S-JSD.
- On nuScenes, Spiral improves over the second-best method by 49.03% (S-FRD), 67.84% (S-FPD), and 46.79% (S-JSD).
- Generative data augmentation at 1% real labels raises SemanticKITTI mIoU from 37.76 to 47.41 with Spiral samples versus 44.16 with R2DM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add closed-loop semantic feedback during inference instead of using open-loop inference throughout (ablation) | Spiral · SemanticKITTI | S-FRD: 395.64 -> 382.87 | Helped |
| add closed-loop semantic feedback during inference instead of using open-loop inference throughout (ablation) | Spiral · SemanticKITTI | S-FPD: 166.95 -> 153.61 | Helped |
| set confidence threshold δ to 0.3 rather than the default 0.8 (ablation) | Spiral closed-loop inference · SemanticKITTI | S-FRD: 382.87 -> 444.31 | Hurt |
| set confidence threshold δ to 0.5 rather than the default 0.8 (ablation) | Spiral closed-loop inference · SemanticKITTI | S-FRD: 382.87 -> 396.71 | Hurt |
| set confidence threshold δ to 0.6 rather than the default 0.8 (ablation) | Spiral closed-loop inference · SemanticKITTI | S-FRD: 382.87 -> 392.36 | Hurt |
| set confidence threshold δ to 0.7 rather than the default 0.8 (ablation) | Spiral closed-loop inference · SemanticKITTI | S-FRD: 382.87 -> 385.05 | Hurt |
| set confidence threshold δ to 0.9 rather than the default 0.8 (ablation) | Spiral closed-loop inference · SemanticKITTI | S-FRD: 382.87 -> 384.42 | Hurt |
| use Spiral-generated scenes rather than R2DM-generated scenes for augmentation | SPVCNN++ segmentation training · SemanticKITTI, 1% real labels | mIoU: 44.16 -> 47.41 | Helped |
| use Spiral-generated scenes rather than R2DM-generated scenes for augmentation | SPVCNN++ segmentation training · SemanticKITTI, 10% real labels | mIoU: 60.62 -> 61.14 | Helped |
| use Spiral-generated scenes rather than R2DM-generated scenes for augmentation | SPVCNN++ segmentation training · SemanticKITTI, 20% real labels | mIoU: 61.21 -> 62.35 | Helped |
| add Spiral-generated scenes to segmentation training | SPVCNN++ · Robo3D fog, 1% real labels | mIoU: 32.07 -> 44.06 | Helped |
| add Spiral-generated scenes to segmentation training | SPVCNN++ · Robo3D wet-ground, 1% real labels | mIoU: 34.26 -> 44.19 | Helped |
Surprising findings
- SPVCNN++, the stronger segmentation model on real data, performs worse than RangeNet++ on scenes generated by LiDARGen and LiDM, suggesting sensitivity to generation noise.
- Closed-loop inference with a low confidence threshold (0.3) performs worse than open-loop inference on semantic-aware metrics.
Limitations
- Generated scenes still contain semantic and geometric artifacts, as shown in qualitative results.
- Evaluation of semantic quality relies on pretrained segmenters, since ground-truth labels for generated scenes are unavailable.
- Real-time performance is still a practical concern (5.7 s per sample on A6000).
Experimental setup
Datasets: SemanticKITTI nuScenes Robo3D
Benchmarks: SemanticKITTI nuScenes Robo3D fog Robo3D wet-ground
Metrics: S-FRD S-FPD S-MMD S-JSD mIoU
Baselines: LiDARGen LiDM R2DM RangeNet++ SPVCNN++
Builds on: LiDARGen LiDM R2DM DDPM Efficient U-Net
Closest papers by meaning
- Towards Realistic Scene Generation with LiDAR Diffusion Models (CVPR 2024 · similarity 0.88)
- LiFi: LiDAR Generation from Multi-View Images via Geometric and Semantic Collaborative Guidance (NeurIPS 2026 · similarity 0.88)
- PC-BEV: An Efficient Polar-Cartesian BEV Fusion Framework for LiDAR Semantic Segmentation (AAAI 2025 · similarity 0.87)
- TREND: Unsupervised 3D Representation Learning via Temporal Forecasting for LiDAR Perception (NeurIPS 2025 · similarity 0.87)
- LiDAR-LLM: Exploring the Potential of Large Language Models for 3D LiDAR Understanding (AAAI 2025 · similarity 0.86)
- SemCity: Semantic Scene Generation with Triplane Diffusion (CVPR 2024 · similarity 0.86)
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.