Research map / 3D vision / Face and shape
Bayesian Diffusion Models for 3D Shape Reconstruction
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 Bayesian Diffusion Models (BDM), which couple a prior point-cloud diffusion model with a single-view reconstruction diffusion model via step-wise fusion (learned merging or training-free blending) to improve 3D shape reconstruction under limited paired data.
Problem
Single-view 3D shape reconstruction trained on paired image-point cloud data is data-hungry; standalone 3D shape (label-only) datasets are abundant but classical Bayesian priors are hard to combine with deep data-driven models.
Approach
Replace explicit prior and likelihood in Bayesian inference (Langevin dynamics) with two jointly running diffusion processes, a top-down prior diffusion model and a bottom-up reconstruction diffusion model, fused at intermediate denoising steps either by a learned merge of decoder features (BDM-M) or by training-free point-level blending (BDM-B).
Key results
- On ShapeNet-R2N2 chair with 10% data, BDM-B reduces CCD-3DR Chamfer Distance from 89.79 to 79.26 and raises F1 from 0.418 to 0.441.
- On Pix3D, BDM-B lowers PC2 table CD from 202.77 to 186.46 and raises F1 from 0.397 to 0.429.
- Human evaluation (16 raters, 1-3 scale) averages 2.35 for BDM-M versus 1.67 for CCD-3DR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Blend prior and reconstruction point clouds | CCD-3DR · ShapeNet-R2N2 Chair 10% | CD: 89.79 -> 79.26 | Helped |
| Blend prior and reconstruction point clouds | CCD-3DR · ShapeNet-R2N2 Chair 100% | CD: 58.47 -> 56.78 | Helped |
| Blend prior and reconstruction point clouds | CCD-3DR · ShapeNet-R2N2 Airplane 10% | CD: 81.29 -> 77.34 | Helped |
| Blend prior and reconstruction point clouds | CCD-3DR · ShapeNet-R2N2 Airplane 100% | CD: 62.77 -> 56.96 | Helped |
| Blend prior and reconstruction point clouds | CCD-3DR · ShapeNet-R2N2 Car 10% | CD: 63.13 -> 59.05 | Helped |
| Blend prior and reconstruction point clouds | CCD-3DR · ShapeNet-R2N2 Car 100% | CD: 61.88 -> 66.90 | Hurt |
| Blend prior and reconstruction point clouds | CCD-3DR · Pix3D Chair | CD: 111.42 -> 111.12 | Helped |
| Blend prior and reconstruction point clouds | CCD-3DR · Pix3D Sofa | CD: 44.91 -> 44.51 | Helped |
| Blend prior and reconstruction point clouds | CCD-3DR · Pix3D Table | CD: 196.28 -> 194.91 | Helped |
| Integrate prior during early and late denoising stages (ablation) | BDM-B · ShapeNet Chair | Chamfer Distance: 89.79 -> 79.26 | Helped |
| Integrate prior during the middle denoising stage (ablation) | BDM-B · ShapeNet Chair | F-Score@0.01: 0.418 -> 0.395 | Hurt |
| Increase prior integration duration from 0 to 16 steps (ablation) | BDM-B · ShapeNet Chair | Chamfer Distance: 89.79 -> 79.26 | Helped |
Surprising findings
- Integrating the prior in the middle denoising stage degraded performance relative to baseline on F1.
- Increasing the prior integration ratio beyond 50% sharply degrades reconstruction (CD rises to 256.13 at 100%).
Limitations
- BDM requires both the prior and data-driven processes to be diffusion processes.
- BDM-B relies on explicit point cloud representation and may not transfer to implicit representations.
Experimental setup
Datasets: ShapeNet-R2N2 Pix3D
Benchmarks: ShapeNet-R2N2 Pix3D
Metrics: Chamfer Distance F-Score@0.01
Baselines: PC2 CCD-3DR CFG
Builds on: PVD PC2 CCD-3DR Stochastic gradient Langevin dynamics Classifier-free guidance
Closest papers by meaning
- Adaptive 3D Reconstruction via Diffusion Priors and Forward Curvature-Matching Likelihood Updates (NeurIPS 2025 · similarity 0.87)
- A Mixture-Based Framework for Guiding Diffusion Models (ICML 2025 · similarity 0.87)
- DiffusionSfM: Predicting Structure and Motion via Ray Origin and Endpoint Diffusion (CVPR 2025 · similarity 0.86)
- Divide-and-Conquer Posterior Sampling for Denoising Diffusion priors (NeurIPS 2024 · similarity 0.86)
- Motion2VecSets: 4D Latent Vector Set Diffusion for Non-rigid Shape Reconstruction and Tracking (CVPR 2024 · similarity 0.86)
- DiffusionBlend: Learning 3D Image Prior through Position-aware Diffusion Score Blending for 3D Computed Tomography Reconstruction (NeurIPS 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.