Abstract
Full-reference (FR) point cloud quality assessment (PCQA) has achieved impressive progress in recent years. However, in many cases, obtaining the reference point cloud is difficult, so the no-reference (NR) methods have become a research hotspot. Few researches about NR objective quality metrics are conducted due to the lack of a large-scale subjective point cloud dataset. Besides, the distinctive property of the point cloud format makes it infeasible to apply blind image quality assessment (IQA) methods directly to predict the quality scores of point clouds. In this paper, we establish a large-scale PCQA dataset, which includes 104 reference point clouds and more than 24,000 distorted point clouds. In the established dataset, each reference point cloud is augmented with 33 types of impairments at 7 different distortion levels. Besides, inspired by the hierarchical perception system and considering the intrinsic attributes of point clouds, an end-to-end sparse convolutional neural network (CNN) is designed to accurately estimate the subjective quality. The results demonstrate that the proposed network has reliable performance.
Introduction
Point clouds serve as the foundation for immersive media, autonomous driving, and robotics. However, these 3D representations are susceptible to various distortions during acquisition, compression, and transmission. Measuring the perceived quality of these models is critical for ensuring a high-quality user experience.
While Full-Reference (FR) metrics exist, they require an original, undistorted point cloud for comparison—a luxury rarely available in real-time streaming or live capture scenarios. No-Reference (NR) or "blind" metrics are more practical but have historically faced three major hurdles: extreme data scarcity, the high cost of human annotation (Mean Opinion Scores), and the computational inefficiency of applying standard 2D "dense" convolutions to sparse 3D data.
The LS-PCQA Dataset
To address the data gap, the authors introduced LS-PCQA, the largest dataset of its kind. It contains 104 reference point clouds and 22,568 distorted samples across 31 impairment types (e.g., Gaussian noise, local missing points, and compression losses) at 7 distortion levels.
To annotate this massive volume of data, a "Pseudo MOS" strategy was employed. The researchers conducted a controlled subjective study on a 1,240-sample subset to identify which FR metrics best correlated with human judgment for specific distortions. These metrics were then used to automatically label the remaining tens of thousands of samples, effectively bootstrapping a large-scale training set.
ResSCNN: Sparse Learning for 3D Quality
The proposed ResSCNN is an end-to-end network that operates directly on 3D coordinates and color attributes. Key architectural features include:
- Sparse Convolutions: Utilizing the Minkowski Engine, the network calculates only where points exist, drastically reducing memory overhead.
- Hierarchical Features: The model extracts features at multiple depths to capture both fine-grained texture/color details and broader structural information.
- Residual Connections: These help retain critical information during the pooling process, mimicking the multi-scale nature of human quality perception.
Key Results and Discussion
ResSCNN achieved state-of-the-art performance among NR-PCQA metrics across multiple benchmarks. On the SJTU-PCQA dataset, it reached an SROCC of 0.81, far exceeding previous NR models like PQA-Net (0.23). Notably, pre-training on the large-scale LS-PCQA dataset improved performance on smaller benchmarks by up to 40%, confirming that data scarcity was a primary bottleneck in the field.
Despite its high accuracy, the model remains efficient with only 1.2M parameters. However, the authors noted that NR metrics still face challenges with color-only distortions (like contrast changes) where the lack of a reference makes it difficult to distinguish between natural lighting and artificial distortion.
Conclusion
The introduction of the LS-PCQA dataset and the ResSCNN architecture represents a significant step forward for blind 3D quality assessment. By leveraging sparse convolutions and large-scale data bootstrapping, this approach provides a robust framework for evaluating 3D media in real-world applications where reference models are unavailable.
Summary updated