UNIQ: Unsupervised Point Cloud Quality Assessment Via Natural Statistics Modeling

Effective Point Cloud Quality Assessment (PCQA) is vital in optimizing the Quality of Experience at the system level. However, existing PCQA approaches often depend on original reference point cloud samples or large datasets with mean opinion score (MOS) labels, limiting their practicality and adaptability across diverse datasets. To address these challenges, we propose a novel UNsupervIsed Quality (UNIQ) index. This index unifies quality prediction in both No-Reference (NR) and Reduced-Reference (RR) modes by leveraging the statistical characteristics of natural 3D scenes, thereby eliminating the need for full-reference data or MOS labels. In NR mode, UNIQ assesses geometry and attribute distortions by measuring the Mahalanobis distance between the statistical parameters of quality-aware geometric and attributive features from pristine and distorted point clouds. In RR mode, it computes the absolute differences between the statistical parameters of the reference and the corresponding distorted point clouds for their respective geometry and attribute components. These geometry and attribute distances are then linearly weighted and normalized by the density factor to generate the final index. Extensive experiments demonstrate that UNIQ achieves competitive performance and exhibits strong generalization to unseen datasets, even surpassing some MOS-supervised methods. This makes UNIQ a promising solution for practical applications, such as point cloud streaming, where MOS labels or full-reference data are often unavailable.

• For the first time, UNIQ integrates the NR and RR models within a unified framework through a carefully designed distance calculation approach upon natural statistics of quality-aware content features aggregated from the reference and impaired point clouds.
• UNIQ carefully selects and aggregates features based on the stage-wise processing mechanism of the HVS, i.e., fundamental visual features extracted at the initial stage and then density guided normalization at the secondary stage. Notably, to the best of our knowledge, the density factor is introduced for the first time in UNIQ to account for its perceptual impact.
• Compared to existing PCQA approaches, the UNIQ eliminates the need for MOS labels or full-reference original data and exhibits competitive performance and superior generalization to unseen distortions across various datasets.

• Block partition:The input point cloud is divided into smaller processing patches.
• Statistics Aggregation:Quality-aware features are extracted, and their statistical parameters are derived in parallel for all patches.
• Quality Estimation:The statistical parameters are compared either against the global pristine distribution or the patch-wise metadata encapsulated from the original reference sample to derive the index.