Research map / 3D vision / 3D generation
Diffusion-based Missing-view Generation With the Application on Incomplete Multi-view Clustering
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 DMVG, a conditional diffusion (UNet) model that generates missing views from available views, plus a data augmentation strategy (DA-DMVG) for high missing rates, and uses the filled data for incomplete multi-view clustering.
Problem
Incomplete multi-view clustering where views of some samples are missing; most methods only use available views rather than recovering missing views, and prior generative recovery methods (VIGAN, CRA) are unstable or limited to dual views.
Approach
Treat missing-view recovery as a conditional diffusion problem: train one UNet per view to denoise the target view conditioned on all other (possibly missing) views, with random condition dropout (CFG-style) and augmentation that creates extra training samples with more missing views, then fill missing views before any MVC method.
Key results
- On Multi-Modal CelebA-HQ with 10% missing views, DMVG reaches RMSE 0.24 / PSNR 12.34 for RGB versus VIGAN's 0.25 / 12.10.
- On Carl with 10% missing views, DMVG reduces RMSE for the classic view to 0.11 versus CRA's 0.16 and DCP's 0.60.
- On Caltech7 with 10% missing views, DAD+OPIMC improves ACC from 49.24% to 57.41% and DAD+PIC from 58.82% to 65.70% (ACC).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use conditional diffusion for missing-view generation instead of VIGAN | Multi-Modal CelebA-HQ · Multi-Modal CelebA-HQ, 10% missing, RGB | RMSE: 0.25 -> 0.24 | Helped |
| Use conditional diffusion for missing-view generation instead of CRA | Carl · Carl, 10% missing, classic | RMSE: 0.16 -> 0.11 | Helped |
| Use conditional diffusion for missing-view generation instead of CRA | NH-face · NH-face, 10% missing, gabor | RMSE: 0.19 -> 0.04 | Helped |
| Add data augmentation to DMVG (ablation) | DMVG on NH-face · NH-face, 50% missing, grey | RMSE: 0.17 -> 0.10 | Helped |
| Add data augmentation to DMVG (ablation) | DMVG on NH-face · NH-face, 50% missing, gabor | RMSE: 0.08 -> 0.08 | No clear effect |
| Train without data augmentation for 1000 rather than 500 epochs (ablation) | DMVG on NH-face · NH-face, 50% missing, grey | RMSE: 0.17 -> 0.12 | Helped |
| Fill missing views with DA-DMVG before clustering | PIMVC on Caltech7 · Caltech7, 30% missing | ACC: 67.35 -> 68.40 | Helped |
| Fill missing views with DA-DMVG before clustering | PIMVC on Caltech7 · Caltech7, 10% missing | ACC: 67.50 -> 67.88 | Helped |
Surprising findings
- Filling missing views does not always improve clustering; the paper notes generated noise can hurt, e.g., DAD+CCR-MVSC and DAD+PIMVC gains are small or mixed at some missing rates.
Limitations
- Generated views can be blurry (e.g., on Carl), and filling views may introduce noise that degrades clustering for some methods.
- Requires training a separate diffusion model per view, and AE compression is needed for high-dimensional image data.
- Diffusion-based generation is costly (T=1000 steps) and the paper does not report compute.
Experimental setup
Datasets: Caltech7 NH-face Multi-Modal CelebA-HQ Carl
Benchmarks: Multi-Modal CelebA-HQ Carl NH-face Caltech7
Metrics: RMSE NMSE PSNR ACC NMI PUR
Baselines: VIGAN CRA DCP BSV Concat MultiNMF CCR-MVSC DAIMC OPIMC EEIMVC PIC IMSR
Builds on: DDPM Classifier-free diffusion guidance VIGAN CRA DCP
Closest papers by meaning
- Incomplete Multi-view Clustering via Diffusion Contrastive Generation (AAAI 2025 · similarity 0.94)
- Incomplete Contrastive Multi-View Clustering with High-Confidence Guiding (AAAI 2024 · similarity 0.91)
- Adaptive Feature Imputation with Latent Graph for Deep Incomplete Multi-View Clustering (AAAI 2024 · similarity 0.90)
- Sample-Level Cross-View Similarity Learning for Incomplete Multi-View Clustering (AAAI 2024 · similarity 0.90)
- Towards Resource-friendly, Extensible and Stable Incomplete Multi-view Clustering (ICML 2024 · similarity 0.89)
- Deep Variational Incomplete Multi-View Clustering: Exploring Shared Clustering Structures (AAAI 2024 · similarity 0.89)
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.