TL;DR
This paper introduces RBFIM, a novel method for assessing the perceived quality of compressed 3D point clouds. By using Radial Basis Function (RBF) interpolation to create a continuous feature representation of the distorted point cloud, RBFIM establishes precise one-to-one correspondences between original and compressed points, significantly improving quality assessment accuracy. This approach offers a more robust and efficient alternative to traditional nearest-neighbor methods, especially for point clouds with varying densities and positions after compression.
What Problem Does This Paper Address?
The increasing use of 3D point clouds in applications like augmented reality and autonomous navigation necessitates efficient compression techniques. However, a critical challenge in point cloud compression (PCC) is accurately evaluating the perceived quality of the compressed output. Existing methods often struggle with establishing reliable correspondences between points in the original and distorted point clouds. Traditional approaches, like searching for nearest neighbors, fail when point clouds have significantly different densities or point positions after compression, leading to inaccurate quality assessments. This paper aims to bridge this gap by proposing a method that can accurately capture human perceptual features despite these compression-induced distortions.
Methods
The core of the proposed method, RBFIM (Radial Basis Function Interpolation Metric), lies in its novel approach to establishing feature correspondences. Instead of relying on direct nearest-neighbor searches, RBFIM first converts the discrete features of the distorted point cloud into a continuous feature function using Radial Basis Function (RBF) interpolation. This function essentially models the spatial distribution of features in the distorted point cloud.
The process involves:
- Preprocessing: Normalizing both the original and distorted point clouds to handle variations in geometry scales.
- Feature Function Calculation:
- Local and Global Feature Functions: RBF interpolation is used to create a continuous feature function for the distorted point cloud. This involves calculating local feature functions for subdomains of the point cloud (using an adaptive octree partitioning method for efficiency) and then combining them using non-negative blending functions to form a global feature function. Luminance is used as the primary feature, though geometry features like curvature can also be employed.
- Reference Points: The original point cloud's points are used as reference points.
- Bijective Feature Set Creation: By substituting the coordinates of the reference points (from the original point cloud) into the continuous feature function of the distorted point cloud, RBFIM generates bijective sets of point features. This ensures a one-to-one mapping between features of the original and distorted point clouds, overcoming the many-to-one correspondence issue common in nearest-neighbor methods.
- Distortion Calculation: The perceptual distortion is then calculated by comparing the feature differences between these precisely corresponding points. The paper uses Mean Squared Error (MSE) on luminance values within octree nodes as the distortion measure, which is then converted to a perceptual quality score.
Datasets
The paper evaluates RBFIM on several publicly available point cloud datasets: NWPU-PCQD, WPC, ICIP2020, IRPC, and M-PCCD. These datasets are used to generate both G-PCC (Geometry-based Point Cloud Compression) and V-PCC (Video-based Point Cloud Compression) distortions. For comprehensive analysis, these datasets are amalgamated into "ALL G-PCC," "ALL V-PCC," and a unified "ALL dataset."
Key Insights and Ideas
- Continuous Feature Representation: The key innovation is transforming discrete point features into a continuous function. This allows for interpolation and accurate feature extraction at arbitrary locations, crucial for handling varying point densities.
- Bijective Correspondence: Establishing a one-to-one mapping between original and distorted point features is critical for accurate quality assessment. RBFIM achieves this by using the original point cloud's geometry as input to the distorted point cloud's feature function.
- Efficiency through Adaptive Partitioning: To manage the computational cost of RBF interpolation on large point clouds, an adaptive octree partitioning method is employed to compute local feature functions, significantly improving efficiency.
- Robustness to Compression Artifacts: The method is designed to be robust to significant changes in point number and position, which are common artifacts in PCC.
Key Results
RBFIM demonstrates superior performance compared to 15 other classic and state-of-the-art PCQA metrics across multiple datasets and distortion types.
- Overall Performance: RBFIM consistently ranks among the top two performing metrics, particularly excelling in G-PCC distortions.
- Quantitative Metrics: On the combined "ALL" dataset, RBFIM achieved the best performance in terms of PLCC and RMSE, and second-best in SROCC and KROCC.
- Average Rankings: Across all datasets, RBFIM achieved the highest average rankings for PLCC, SROCC, KROCC, and RMSE.
- Computational Efficiency: RBFIM exhibits competitive running times, often faster than multi-feature metrics, due to its simplified correspondence establishment.
Limitations and Future Work
- Feature Choice: While luminance is shown to be effective, the paper notes that other features (like curvature) can also be used, though they might add complexity or have limitations when only attributes are lost. The optimal combination of features for various scenarios is an area for future investigation.
- M-PCCD Dataset: Performance on the M-PCCD dataset was slightly lower than on others, suggesting that unique distortion types or features might require more nuanced representations, potentially involving multi-scale feature assessment.
- Future Work: The authors plan to explore more features for RBFIM to establish a more universally applicable metric and investigate combinations of features to further enhance performance.
Summary updated