No-Reference Point Cloud QA

Updated 11 October 2025

The paper introduces NR-PCQA methods that predict perceptual quality of 3D point clouds without a pristine reference by leveraging large annotated datasets and pseudo-labeling techniques.
It employs diverse methodologies including direct 3D, patch-wise, projection-based, multimodal, contrastive pre-training, and graph-based models to capture complex distortions.
Cross-dataset evaluations show that NR-PCQA models achieve high correlation with human scores, ensuring robust quality prediction for immersive and real-time applications.

No-Reference Point Cloud Quality Assessment (NR-PCQA) encompasses methodologies designed to objectively predict the perceptual quality of 3D point clouds in the absence of a pristine reference. This is distinguished from full-reference (FR) PCQA by its capacity to operate in real-world scenarios where only the potentially distorted or compressed point cloud is present. NR-PCQA is essential for immersive media applications, real-time streaming, adaptive 3D transmission, and autonomous systems, where ground-truth models are not available.

1. Dataset Construction and Annotation Strategies

Progress in NR-PCQA has been accelerated by the curation of large, diversified datasets and the invention of robust pseudo-labeling techniques. The LS-PCQA dataset (Liu et al., 2020) fundamentally shaped the field by providing over 22,000 distorted samples of 104 reference point clouds subjected to 31 impairment types at multiple severity levels. Both geometry-related (down-sampling, local missing, geometric shifts) and photometric (color noise, contrast, compression loss) artifacts are included, yielding comprehensive coverage of typical degradations. Because exhaustive subjectively scored datasets are infeasible at this scale, LS-PCQA uses a two-phase approach: a subset is annotated with Mean Opinion Scores (MOS) from human raters, then the best FR metric per distortion (judged by SROCC with subjective scores) is used with a logistic normalization mapping to extrapolate "pseudo MOS" labels across the full dataset.

Such strategies facilitate supervised learning on a scale previously unattainable, enabling robust deep NR models and generalization studies. Additional databases, such as SJTU-PCQA, WPC, BASICS, WPC5.0, and WPC6.0 (Long et al., 9 Oct 2024, Duan et al., 9 Oct 2024), provide varied content, distortion regimes, and encoding pipelines, anchoring cross-dataset evaluation and broadening applicability.

2. NR-PCQA Methodological Taxonomy

2.1 Direct 3D and Patch-wise Architectures

A prominent class of NR-PCQA methods evaluates the raw 3D point set directly. For example, ResSCNN (Liu et al., 2020) is a sparse convolutional neural network that extracts multiscale features from the entire point cloud, leveraging hierarchical perception analogies from human vision. By operating on a sparse tensor representation (coordinates and attributes), and using submanifold convolutions, it encodes high-frequency color and geometric detail in shallow layers and global structure in deeper layers. Patch-based models, such as COPP-Net (Cheng et al., 2023), first segment the cloud into patches, predict patch-level quality with feature fusion (texture/structure), and aggregate predictions with learned correlation weights—better reflecting the unequal contribution of different regions to overall perceptual quality.

2.2 Projection and Video-based Approaches

Projecting the point cloud onto multi-view 2D images is computationally efficient and leverages established 2D vision backbones (ConvNeXt V2, Swin Transformer (Zhang et al., 2023)). Cube-like projection (six views), circular camera orbits (Fan et al., 2022), and multi-perspective rendering encapsulate both global form and detail. Some approaches, such as the video-based metric (Fan et al., 2022), employ dynamic virtual camera motion to capture temporal cues that approximate human exploratory observation.

Emergent frameworks adopt multi-modal learning, fusing features from 3D geometry branches (e.g., PointNet++ for patch geometry) and 2D texture branches (e.g., ResNet for image projections), to better exploit the complementary sensitivity of each modality (Zhang et al., 2022). The use of symmetric cross-modal attention mechanisms (Zhang et al., 2022) allows texture and geometric features to guide one another, improving the discriminability of subtle artifacts.

2.4 Contrastive and Self-Supervised Pre-training

To address label scarcity and generalization, self-supervised and contrastive pre-training regimes have become critical. Masked autoencoder frameworks such as PAME (Shan et al., 15 Mar 2024) reconstruct masked patches of projected images, learning to disentangle content and distortion features without labels. Contrastive pre-training (CoPA (Shan et al., 15 Mar 2024)) creates mixed "anchor" images by patch-wise mixing between different distortion types, using a quality-aware contrastive loss to emphasize both inter-distortion and inter-content discrimination.

2.5 Probabilistic and Explainable Models

Traditional architectures regress a deterministic quality score, but recent works explicitly model the stochasticity of human ratings by formulating PCQA as conditional probabilistic inference. For example, a conditional variational autoencoder (CVAE) (Fan et al., 17 Jan 2024) produces multiple samples (virtual "subjective judgments") and integrates them for a final prediction, resulting in higher robustness and cross-dataset consistency. Approaches such as CLIP-PCQA (Liu et al., 17 Jan 2025) align vision-language representations by retrieving the closest textual description (e.g., "good," "poor") and predict the full opinion score distribution, further capturing subjective uncertainty.

2.6 Graph and Attention-based Processing

Perceptual Clustering Weighted Graph (PCW-Graph) (Laazoufi et al., 4 Jun 2025) methods segment the cloud into clusters via K-means in a space defined by local color (LAB), curvature, and saliency. Clusters serve as graph nodes, with edge weights reflecting both geometric adjacency and perceptual similarity. Graph transformer attention fuses these region-level features, explicitly modeling spatial relationships and perceptual importance.

3. Notable Model Formulations and Technical Details

The diversity of NR-PCQA models is reflected in their mathematical formalism and feature selection. Key recurrent elements include:

3D-Natural Scene Statistics (3D-NSS): Extraction of handcrafted statistical features (curvature, anisotropy, entropy) from both geometry and LAB color domains, with parametric modeling (\emph{e.g.}, GGD, AGGD, Gamma distributions) (Zhang et al., 2021).
Attention mechanisms: Multi-head self- and cross-attention layers (as in transformers) learn semantic affinities and fuse color-geometry cues on points/patches or clusters (Tliba et al., 2023, Laazoufi et al., 4 Jun 2025).
Bitstream-layer metrics: Analytical models such as streamPCQ-TL (Long et al., 9 Oct 2024) and streamPCQ-OR (Duan et al., 9 Oct 2024) exploit parameters directly from the compressed bitstream (e.g., TBPP, TQP, PQS, tNSL) to compute content-adaptive distortion factors, avoiding the need for full decoding and affording ultra-efficient quality prediction.
Mutual Information Minimization: DisPA (Shan et al., 12 Nov 2024) learns separate representations for content and distortion, minimizing the estimated mutual information between them to ensure disentanglement, resulting in superior adaptation to complex and varied distortion regimes.
Multimodal Large Model Fusion: PIT-QMM (Gupta et al., 9 Oct 2025) combines point cloud tokens, image projections, and textual context as input to a large language/vision model, supporting both quality prediction and distortion localization.

4. Performance Evaluation and Comparative Results

Experimental assessment is standardized across PLCC, SROCC, KRCC, and RMSE metrics. NR-PCQA models such as ResSCNN (Liu et al., 2020), MM-PCQA (Zhang et al., 2022), COPP-Net (Cheng et al., 2023), and others routinely bridge or close the performance gap with FR metrics, and in some instances exceed FR performance on challenging test samples. For instance, ResSCNN achieves PLCC/SROCC values of ~0.60–0.62 on LS-PCQA (Liu et al., 2020), while COPP-Net and MM-PCQA report PLCC exceeding 0.93 and SRCC near 0.91 on challenging WPC and SJTU-PCQA datasets (Cheng et al., 2023, Zhang et al., 2022). Cross-dataset evaluations, as emphasized in recent works (Fan et al., 17 Jan 2024, Shan et al., 12 Nov 2024, Gupta et al., 9 Oct 2025), demonstrate generalization—essential for practical deployment where test-time distortions and content diverge from training distributions.

Methodological ablation studies consistently confirm the benefit of hierarchical feature extraction, content–distortion disentanglement, region-wise weighting, and cross-modal attention mechanisms.

5. Open Challenges, Future Directions, and Research Implications

Open problems in NR-PCQA research include:

Generalization to Unseen Distortions: Ensuring models are not overfitted to known artifact types and dataset-specific biases. Pre-training with vast unlabeled data and explicit content–distortion separation remain promising.
Label Scarcity and Subjective Uncertainty: Techniques such as probabilistic modeling of opinion score distributions (Fan et al., 17 Jan 2024, Liu et al., 17 Jan 2025) and language-driven retrieval-based inference mechanisms address the intrinsic stochasticity and ambiguity in subjective assessment.
Multimodality and Explainability: Large multimodal models (LMMs) (Gupta et al., 9 Oct 2025) achieving high accuracy and real-time performance also provide capabilities for distortion localization and explanation. Explainable NR-PCQA is crucial for actionable feedback in professional applications, particularly for industrial 3D asset management.
Bitstream-layer Assessment and Real-time Performance: Bitstream-centric models (Long et al., 9 Oct 2024, Duan et al., 9 Oct 2024) enable ultra-efficient in-network QC, critical for streaming and edge deployment.
Temporal and Dynamic NCQA: Video-based and sequence-aware approaches (Fan et al., 2022) extend NR-PCQA to dynamic 3D content and time-varying scenarios.

6. Accessibility and Resources

Many leading NR-PCQA models, datasets, and evaluation codebases are public:

Resource	Model/Dataset	Link
LS-PCQA	Dataset	http://smt.sjtu.edu.cn
ResSCNN	NR metric (code)	https://github.com/lyp22/ResSCNN
COPP-Net	NR metric (code)	https://github.com/philox12358/COPP-Net
MM-PCQA	NR metric (code)	https://github.com/zzc-1998/MM-PCQA
IT-PCQA	NR metric (code)	https://github.com/Qi-Yangsjtu/IT-PCQA
PAME	NR metric (code)	https://github.com/unike132/pame
CLIP-PCQA	NR metric (code)	https://github.com/cunywang/clip-pcqa
PIT-QMM	NR metric (code)	https://www.github.com/shngt/pit-qmm
streamPCQ-TL	Bitstream metric	https://github.com/qdushl/Waterloo-Point-Cloud-Database-6.0
streamPCQ-OR	Bitstream metric	https://github.com/qdushl/Waterloo-Point-Cloud-Database-5.0

The field of NR-PCQA, propelled by the integration of large datasets, modern neural architectures, efficient statistical and graph-theoretical tools, and language-driven frameworks, is maturing rapidly. These advances provide foundational capabilities for automated, perceptually aligned 3D quality monitoring in an expanding array of real-world and industrial domains.