Research map / Point clouds / Point cloud classification
Learning Disentangled Representations for Perceptual Point Cloud Quality Assessment via Mutual Information Minimization
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 DisPA, a dual-branch NR-PCQA framework that disentangles content and distortion representations by minimizing a variational MI upper bound, with masked-autoencoding content pretraining and mini-patch distortion input.
Problem
No-reference point cloud quality assessment models entangle point cloud content and distortion representations in a single network, and content overfitting (limited content diversity in PCQA datasets) makes predictions deviate when content changes with the same distortion.
Approach
Two branches with different philosophies: a ViT content encoder pretrained by masked autoencoding on rendered multi-view images of distorted-vs-reference point clouds, and a Swin-T distortion encoder on grid mini-patch maps; an MLP-based variational MI upper bound estimator is minimized as an explicit disentanglement regularizer.
Key results
- On SJTU-PCQA, DisPA reaches SROCC 0.908 and PLCC 0.919, versus 0.897 SROCC for the prior best NR method CoPA.
- On WPC, DisPA reaches SROCC 0.788 and PLCC 0.790, and on LS-PCQA SROCC 0.625, the best NR results in the comparison table.
- Ablation on SJTU-PCQA: removing content-aware pretraining drops SROCC from 0.908 to 0.876; replacing MI with cosine similarity gives 0.884.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add content-aware pretraining (ablation) | DisPA · SJTU-PCQA | SROCC: 0.876 -> 0.908 | Helped |
| Add content-aware pretraining (ablation) | DisPA · SJTU-PCQA | PLCC: 0.897 -> 0.919 | Helped |
| Add representation disentanglement (ablation) | DisPA · SJTU-PCQA | SROCC: 0.849 -> 0.908 | Helped |
| Add representation disentanglement (ablation) | DisPA · SJTU-PCQA | PLCC: 0.865 -> 0.919 | Helped |
| Use MI rather than cosine similarity for disentanglement (ablation) | DisPA · SJTU-PCQA | SROCC: 0.884 -> 0.908 | Helped |
| Use MI rather than cosine similarity for disentanglement (ablation) | DisPA · SJTU-PCQA | PLCC: 0.910 -> 0.919 | Helped |
| Use both content-aware and distortion-aware branches instead of the content-aware branch alone (ablation) | DisPA · SJTU-PCQA | SROCC: 0.811 -> 0.908 | Helped |
| Use both content-aware and distortion-aware branches instead of the content-aware branch alone (ablation) | DisPA · SJTU-PCQA | PLCC: 0.832 -> 0.919 | Helped |
| Use both content-aware and distortion-aware branches instead of the distortion-aware branch alone (ablation) | DisPA · SJTU-PCQA | SROCC: 0.763 -> 0.908 | Helped |
| Use both content-aware and distortion-aware branches instead of the distortion-aware branch alone (ablation) | DisPA · SJTU-PCQA | PLCC: 0.778 -> 0.919 | Helped |
| Retain differential ranking loss (ablation) | DisPA · SJTU-PCQA | SROCC: 0.892 -> 0.908 | No clear effect |
| Retain differential ranking loss (ablation) | DisPA · SJTU-PCQA | PLCC: 0.910 -> 0.919 | No clear effect |
Surprising findings
- Cosine-similarity disentanglement achieves fairly competitive performance (SROCC 0.884) compared with the MI-based version (0.908).
Limitations
- Content-aware encoder is pretrained only on LS-PCQA, which limits pretraining scale.
- MI estimation for high-dimensional representations is still relatively costly and may need more efficient estimators.
Experimental setup
Datasets: LS-PCQA SJTU-PCQA WPC ImageNet-1K
Benchmarks: LS-PCQA SJTU-PCQA WPC
Metrics: SROCC PLCC RMSE
Baselines: PQA-Net GPA-Net CoPA MM-PCQA ResSCNN IT-PCQA PCQM MS-GraphSIM
Builds on: grid mini-patch sampling masked autoencoder CLUB (contrastive log-ratio upper bound) differential ranking loss
Closest papers by meaning
- Contrastive Pre-Training with Multi-View Fusion for No-Reference Point Cloud Quality Assessment (CVPR 2024 · similarity 0.90)
- CLIP-PCQA: Exploring Subjective-Aligned Vision-Language Modeling for Point Cloud Quality Assessment (AAAI 2025 · similarity 0.87)
- Neural Point Cloud Diffusion for Disentangled 3D Shape and Appearance Generation (CVPR 2024 · similarity 0.86)
- Depth-Guided Robust and Fast Point Cloud Fusion NeRF for Sparse Input Views (AAAI 2024 · similarity 0.84)
- Digging into Intrinsic Contextual Information for High-fidelity 3D Point Cloud Completion (AAAI 2025 · similarity 0.83)
- LinNet: Linear Network for Efficient Point Cloud Representation Learning (NeurIPS 2024 · similarity 0.83)
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.