Research map / Point clouds / Shape and surface completion
SuperPC: A Single Diffusion Model for Point Cloud Completion, Upsampling, Denoising, and Colorization
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 SuperPC, a single conditional diffusion model that jointly performs point cloud completion, upsampling, denoising (and separately colorization) by conditioning on image and point cloud inputs through raw, local, and global three-level conditions fused via spatial-mix-fusion.
Problem
Point cloud processing tasks (completion, upsampling, denoising, colorization) are usually solved by separate specialized models, which ignores their correlated defects and causes error accumulation and extra compute when chained.
Approach
A three-level-conditioned (raw, local, global) diffusion framework that fuses image and point cloud inputs via a Spatial-Mix-Fusion strategy (dual-spatial early fusion plus attention-based deep fusion), letting one model handle all four tasks.
Key results
- On ShapeNet completion, SuperPC reports DCD 0.387 vs 0.462 for AdaPoinTr; on KITTI-360 denoising DCD 0.327 vs ScoreDenoise 0.369.
- On the combined four-stage task on KITTI-360, SuperPC reaches DCD 0.681 vs 0.725 for the best SOTA chain, EMD 9.58 vs 10.06.
- In sim-to-real generalization (TartanAir to KITTI-360), SuperPC outperforms SOTA on all tasks, e.g. completion DCD 0.789 vs 0.873 for PC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SuperPC instead of specialized methods | ShapeNet completion · ShapeNet | DCD: 0.462 -> 0.387 | Hurt |
| Use SuperPC instead of the strongest baseline | ShapeNet upsampling · ShapeNet | DCD: 0.281 -> 0.293 | Mixed |
| Use SuperPC instead of the strongest baseline | ShapeNet denoising · ShapeNet | DCD: 0.280 -> 0.285 | Mixed |
| Use the unified SuperPC model instead of sequential specialized models | ShapeNet combination · ShapeNet | DCD: 0.509 -> 0.476 | Helped |
| Use SuperPC instead of specialized methods | TartanAir completion · TartanAir | DCD: 0.559 -> 0.538 | No clear effect |
| Use SuperPC instead of specialized methods | TartanAir upsampling · TartanAir | DCD: 0.527 -> 0.492 | Helped |
| Use SuperPC instead of specialized methods | TartanAir denoising · TartanAir | DCD: 0.346 -> 0.298 | Helped |
| Use the unified SuperPC model instead of sequential specialized models | TartanAir combination · TartanAir | DCD: 0.583 -> 0.558 | Helped |
| Use SuperPC instead of specialized methods | KITTI-360 completion · KITTI-360 | DCD: 0.649 -> 0.632 | Helped |
| Use SuperPC instead of specialized methods | KITTI-360 upsampling · KITTI-360 | DCD: 0.597 -> 0.577 | Helped |
| Use SuperPC instead of specialized methods | KITTI-360 denoising · KITTI-360 | DCD: 0.369 -> 0.327 | Helped |
| Use the unified SuperPC model instead of sequential specialized models | KITTI-360 combination · KITTI-360 | DCD: 0.725 -> 0.681 | Helped |
Surprising findings
- On the ShapeNet object-level benchmark, the SuperPC upsampling result is not best; DDPMPU is marginally better, and on ShapeNet some specialized models outperform SuperPC on individual tasks.
Limitations
- Relies on a heavy PointNet++ backbone; efficiency could be improved with a sparse-tensor backbone.
- No real-world object-level benchmark is established.
- Colorization is not evaluated jointly with the other three tasks.
Experimental setup
Datasets: ShapeNet TartanAir KITTI-360
Benchmarks: ShapeNet TartanAir KITTI-360
Metrics: DCD EMD F1
Baselines: AdaPoinTr GradPU DDPMPU ScoreDenoise PD-LTS LiDiff
Builds on: denoising diffusion probabilistic models PointNet++ Feature Pyramid Networks PyTorch3D point rasterizer
Closest papers by meaning
- ComPC: Completing a 3D Point Cloud with 2D Diffusion Priors (ICLR 2025 · similarity 0.89)
- PointInfinity: Resolution-Invariant Point Diffusion Models (CVPR 2024 · similarity 0.88)
- TIGER: Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process (CVPR 2024 · similarity 0.87)
- Adaptive 3D Reconstruction via Diffusion Priors and Forward Curvature-Matching Likelihood Updates (NeurIPS 2025 · similarity 0.87)
- StraightPCF: Straight Point Cloud Filtering (CVPR 2024 · similarity 0.87)
- PCDreamer: Point Cloud Completion Through Multi-view Diffusion Priors (CVPR 2025 · similarity 0.87)
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.