Research map / Point clouds / Shape and surface completion
Scaling Diffusion Models to Real-World 3D LiDAR Scene Completion
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 a point-level DDPM for scene-scale 3D LiDAR scene completion from a single scan, using local point-wise noise, a mean/std noise regularization, and a refinement upsampling network.
Problem
Scene completion from sparse single LiDAR scans; prior diffusion methods either work on range images or voxel grids and do not scale to point-level scene-scale completion.
Approach
Reformulates DDPM noising as a local per-point offset (no normalization of the input scan), conditions denoising on the input scan via classifier-free guidance with a MinkUNet noise predictor, and adds a regularizer pushing predicted noise toward N(0,I).
Key results
- Ours refined achieves mean Chamfer distance 0.375 m and BEV JSD 0.416 m on SemanticKITTI, the best on both metrics versus LMSCNet, LODE, MID and PVD.
- On KITTI-360, Ours refined reaches CD 0.517 m and JSD 0.446 m, beating MID (0.637 / 0.476) without fine-tuning.
- At 0.1 m voxel IoU on SemanticKITTI, Ours refined reaches 24.75% versus 16.98% for MID.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add refinement network (ablation) | LiDiff · SemanticKITTI val | CD [m]: 0.434 -> 0.375 | Helped |
| add refinement network (ablation) | LiDiff · SemanticKITTI val | JSDBEV [m]: 0.444 -> 0.416 | Helped |
| add refinement network (ablation) | LiDiff · KITTI-360 sequence 00 | CD [m]: 0.564 -> 0.517 | Helped |
| add refinement network (ablation) | LiDiff · KITTI-360 sequence 00 | JSDBEV [m]: 0.459 -> 0.446 | Helped |
| add refinement network (ablation) | LiDiff · Our data | CD [m]: 0.518 -> 0.471 | Helped |
| add refinement network (ablation) | LiDiff · Our data | JSDBEV [m]: 0.360 -> 0.341 | Helped |
| add refinement network (ablation) | LiDiff · SemanticKITTI val, grid 0.5 m | IoU [%]: 42.49 -> 40.71 | Hurt |
| add refinement network (ablation) | LiDiff · SemanticKITTI val, grid 0.2 m and 0.1 m | IoU [%]: 33.12 -> 38.92; 11.02 -> 24.75 | Helped |
| add refinement network (ablation) | LiDiff · KITTI-360 sequence 00, grid 0.5 m | IoU [%]: 42.22 -> 40.82 | Hurt |
| add refinement network (ablation) | LiDiff · KITTI-360 sequence 00, grid 0.2 m and 0.1 m | IoU [%]: 32.25 -> 36.08; 10.80 -> 21.34 | Helped |
| add refinement network (ablation) | LiDiff · Our data, grid 0.5 m, 0.2 m, and 0.1 m | IoU [%]: 37.16 -> 38.51; 29.17 -> 40.20; 6.53 -> 17.48 | Helped |
| increase noise regularization weight from r=0.0 to r=1.0, r=3.0, or r=5.0 (ablation) | LiDiff noise predictor · SemanticKITTI val short sequence | CD [m]: 0.529; 0.470; 0.441; 0.445 | Helped |
Surprising findings
- Standard PVD-style point diffusion fails at scene scale (CD 1.256 m on SemanticKITTI, lowest IoU among methods), indicating shape-diffusion methods do not transfer directly to scenes.
- Without regularization the predicted noise has mean around -9 and standard deviation about 526, far from N(0,I).
Limitations
- Cannot perform unconditional generation; requires an input scan as condition.
- SDF-based baselines perform better at coarse grid resolutions (0.5 m IoU) and on denser data.
Experimental setup
Datasets: SemanticKITTI KITTI-360 own Ouster OS-1 128-beam LiDAR data
Benchmarks: SemanticKITTI validation (sequence 08) KITTI-360 sequence 00 own LiDAR data
Metrics: Chamfer distance Jensen-Shannon divergence (BEV) IoU at multiple voxel resolutions
Baselines: LMSCNet LODE Make It Dense (MID) PVD
Builds on: DDPM classifier-free guidance DPM-Solver PVD LION
Closest papers by meaning
- Towards Realistic Scene Generation with LiDAR Diffusion Models (CVPR 2024 · similarity 0.91)
- ComPC: Completing a 3D Point Cloud with 2D Diffusion Priors (ICLR 2025 · similarity 0.88)
- Denoising Diffusion via Image-Based Rendering (ICLR 2024 · similarity 0.87)
- Adaptive 3D Reconstruction via Diffusion Priors and Forward Curvature-Matching Likelihood Updates (NeurIPS 2025 · similarity 0.87)
- Coherent 3D Scene Diffusion From a Single RGB Image (NeurIPS 2024 · similarity 0.87)
- TIGER: Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process (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.