Research map / Point clouds / Point cloud classification
Contrastive Pre-Training with Multi-View Fusion for No-Reference Point Cloud Quality Assessment
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 CoPA, a contrastive pre-training framework for no-reference point cloud quality assessment that builds anchors by patch-mixing projected images of differently distorted point clouds, plus a semantic-guided multi-view fusion module for fine-tuning.
Problem
No-reference point cloud quality assessment (NR-PCQA) suffers from scarce labeled MOS data, leading to poor generalization across datasets.
Approach
Generate contrastive anchors by randomly mixing local 16x16 patches of projected images of the same point cloud under two distortions, treat the source images as positives, and split negatives into distortion-wise and content-wise sets; then fuse multi-view projections with ImageNet semantic features via cross-attention.
Key results
- On LS-PCQA, CoPA+FT reaches SROCC 0.613 and PLCC 0.636, beating the best NR baseline GPA-Net/ResSCNN (0.594/0.624).
- On SJTU-PCQA, CoPA+FT achieves SROCC 0.897 and PLCC 0.913; on WPC, SROCC 0.779 and PLCC 0.785, outperforming all NR methods.
- In cross-dataset evaluation (LS to SJTU), CoPA+FT PLCC 0.644 vs 0.581 for MM-PCQA; ablation shows removing pre-training drops SJTU SROCC from 0.897 to 0.806.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add contrastive pre-training (ablation) | CoPA+FT · SJTU-PCQA | SROCC: 0.806 -> 0.897 | Helped |
| add distortion-wise and content-wise contrast rather than distortion-wise contrast only (ablation) | CoPA+FT · SJTU-PCQA | SROCC: 0.856 -> 0.897 | Helped |
| add distortion-wise contrast rather than content-wise contrast only (ablation) | CoPA+FT · SJTU-PCQA | SROCC: 0.889 -> 0.897 | Helped |
| use semantic-guided multi-view fusion rather than max pooling (ablation) | CoPA+FT · SJTU-PCQA | SROCC: 0.841 -> 0.897 | Helped |
| use semantic-guided multi-view fusion rather than average pooling (ablation) | CoPA+FT · SJTU-PCQA | SROCC: 0.810 -> 0.897 | Helped |
| add ranking loss to MSE-only fine-tuning (ablation) | CoPA+FT · SJTU-PCQA | SROCC: 0.889 -> 0.897 | Helped |
| use CoPA pre-training | PQA-Net · SJTU-PCQA | PLCC: 0.687 -> 0.749 | Helped |
| use CoPA pre-training | PQA-Net · WPC | PLCC: 0.579 -> 0.911 | Helped |
| use CoPA pre-training | MM-PCQA · SJTU-PCQA | PLCC: 0.898 -> 0.910 | Helped |
| use CoPA pre-training | MM-PCQA · WPC | PLCC: 0.774 -> 0.778 | Helped |
| train on LS-PCQA and test on SJTU-PCQA | CoPA+FT · SJTU-PCQA | PLCC: 0.581 -> 0.644 | Helped |
| train on LS-PCQA and test on WPC | CoPA+FT · WPC | PLCC: 0.466 -> 0.516 | Helped |
Surprising findings
- Adding CoPA pre-training to PQA-Net raised WPC PLCC from 0.579 to 0.911 in Table 4, a very large gain that is not discussed in depth.
Limitations
- Relies on ImageNet-pretrained ResNet50 and projection-based rendering, so point cloud information is reduced to 2D views.
- Reported cross-dataset performance remains relatively low (e.g., WPC PLCC 0.533 when trained on SJTU).
- Small labeled benchmarks and 5-fold cross-validation mean results depend on split configuration.
Experimental setup
Datasets: LS-PCQA SJTU-PCQA WPC ImageNet
Benchmarks: LS-PCQA SJTU-PCQA WPC
Metrics: SROCC PLCC RMSE
Baselines: MSE-p2point Hausdorff-p2point MSE-p2plane Hausdorff-p2plane PSNR-yuv PointSSIM PCQM GraphSIM MS-GraphSIM MPED PQA-Net IT-PCQA
Builds on: MoCo SimCLR Re-IQA QPT IT-PCQA
Closest papers by meaning
- No-Reference Point Cloud Quality Assessment via Domain Adaptation (CVPR 2022 · similarity 0.91)
- QD-PCQA: Quality-Aware Domain Adaptation for Point Cloud Quality Assessment (CVPR 2026 · similarity 0.91)
- Learning Disentangled Representations for Perceptual Point Cloud Quality Assessment via Mutual Information Minimization (NeurIPS 2024 · similarity 0.90)
- CLIP-PCQA: Exploring Subjective-Aligned Vision-Language Modeling for Point Cloud Quality Assessment (AAAI 2025 · similarity 0.89)
- PCDreamer: Point Cloud Completion Through Multi-view Diffusion Priors (CVPR 2025 · similarity 0.84)
- Digging into Intrinsic Contextual Information for High-fidelity 3D Point Cloud Completion (AAAI 2025 · similarity 0.84)
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.