Abstract
The evolution of compression and enhancement algorithms necessitates an accurate quality assessment for point clouds. Previous works consistently regard point cloud quality assessment (PCQA) as a MOS regression problem and devise a deterministic mapping, ignoring the stochasticity in generating MOS from subjective tests. Besides, the viewpoint switching of 3D point clouds in subjective tests reinforces the judging stochasticity of different subjects compared with traditional images. This work presents the first probabilistic architecture for no-reference PCQA, motivated by the labeling process of existing datasets. The proposed method can model the quality judging stochasticity of subjects through a tailored conditional variational autoencoder (CVAE) and produces multiple intermediate quality ratings. These intermediate ratings simulate the judgments from different subjects and are then integrated into an accurate quality prediction, mimicking the generation process of a ground truth MOS. Specifically, our method incorporates a Prior Module, a Posterior Module, and a Quality Rating Generator, where the former two modules are introduced to model the judging stochasticity in subjective tests, while the latter is developed to generate diverse quality ratings. Extensive experiments indicate that our approach outperforms previous cutting-edge methods by a large margin and exhibits gratifying cross-dataset robustness.
Introduction
Point clouds have become indispensable for various 3D applications, ranging from autonomous driving to cultural heritage preservation. However, the processes of acquisition, compression, and transmission inevitably introduce distortions that degrade visual quality. Assessing this quality without a pristine reference—the No-Reference (NR) problem—is a significant challenge.
Traditional NR-PCQA methods typically treat quality assessment as a deterministic regression task, mapping a point cloud directly to a single Mean Opinion Score (MOS). This approach overlooks the inherent "stochasticity" of human perception. In subjective tests, different individuals provide varying scores for the same content due to personal sensitivity, viewing conditions, and the interactive nature of 3D viewpoint switching. By regressing only to the average (MOS), existing models ignore this rich "judgment noise," which can lead to overconfident predictions and poor generalization across different datasets.
Methods
To address these limitations, we propose a stochasticity-aware NR-PCQA framework based on a Conditional Variational Autoencoder (CVAE). This architecture mimics the actual process of subjective testing by modeling the distribution of human judgments rather than just the mean.
The framework consists of three primary components:
- Prior & Posterior Modules: These modules model the distribution of human judgments in a latent space. The Prior Module estimates the distribution based on point cloud projections. During training, a Posterior Module utilizes the ground-truth MOS to refine this distribution, ensuring the latent space captures meaningful uncertainty.
- Quality Rating Generator (QRG): This component combines deterministic features (representing physical distortions) with stochastic features sampled from the latent distribution. It generates multiple intermediate quality ratings, effectively simulating the diverse judgments of multiple "subjects."
- RGB-D Projections: Point clouds are projected into 2D RGB-D (color + depth) images. We employ random viewpoint selection during training to simulate the interactive viewing behavior of human subjects, which acts as a powerful regularization technique.
To ensure stability, we utilize KL annealing to prevent posterior collapse and a Gaussian Stochastic Neural Network (GSNN) branch to maintain performance during inference when ground-truth labels are unavailable.
Experimental Results
The model was rigorously evaluated on three major benchmarks: SJTU-PCQA, WPC, and WPC2.0. We employed k-fold cross-validation and cross-dataset testing to evaluate robustness.
As shown in the results, our method achieves state-of-the-art performance across all metrics (SRCC, PLCC, KRCC). Notably, on the WPC2.0 dataset, it achieved an SRCC of 0.8742, a substantial improvement over the previous best of 0.8023. Furthermore, our NR model outperformed several Full-Reference (FR) methods, such as GraphSIM and PointSSIM, despite not having access to the original undistorted data.
Cross-dataset experiments further highlighted the model's superior generalization. When training on WPC and testing on SJTU, our approach showed a 14.4% improvement in SRCC over the next best competitor, demonstrating that modeling judgment uncertainty leads to more robust feature representations.
Conclusion
By embracing the stochastic nature of human perception, this work moves beyond deterministic regression for point cloud quality assessment. Our CVAE-based architecture effectively simulates the subjective testing process, resulting in higher accuracy and better generalization. Future research will explore applying these probabilistic principles directly to 3D representations and leveraging raw individual subject scores to further refine the modeling of human perceptual biases.
Summary updated