Inferring Point Cloud Quality via Graph Similarity.

Area of Science:

• Computer Vision
• 3D Data Processing
• Perceptual Quality Assessment

Background:

• Objective quality estimation is crucial for 3D media applications.
• Existing metrics for 2D media lack effectiveness for unstructured 3D point clouds.
• Human vision prioritizes structural details over individual point intensities.

Purpose of the Study:

• To propose GraphSIM, a novel metric for objective quality estimation of 3D point clouds.
• To accurately predict human perception of point clouds with superimposed geometry and color impairments.
• To address the lack of effective quality assessment metrics for 3D point cloud data.

Main Methods:

• Utilizing graph signal gradients to quantify point cloud distortions.
• Extracting geometric keypoints to form object skeletons for local graph construction.
• Computing color gradient moments for local significance similarity and overall pooling for a final similarity index.

Main Results:

• GraphSIM demonstrates state-of-the-art performance across various impairments (re-sampling, compression, noise).
• Achieved noticeable gains in predicting subjective Mean Opinion Scores (MOS) compared to existing metrics.
• Validated on two independent, large-scale point cloud assessment datasets.

Conclusions:

• GraphSIM offers a robust and accurate method for 3D point cloud quality assessment.
• The metric shows generalizability across different scenarios with parameter adjustments.
• Proposed models and materials are publicly available for further research and application.