Point Cloud Quality Assessment Using a One-Dimensional Model Based on the Convolutional Neural Network.

Area of Science:

• Computer Vision
• 3D Data Processing
• Machine Learning

Background:

• 3D modeling advancements impact VR, diagnosis, and architecture.
• Distortions from simplification/compression degrade 3D point cloud quality.
• Objective quality assessment methods are crucial for distorted 3D data.

Purpose of the Study:

• To develop a novel no-reference (NR) deep learning methodology for 3D point cloud quality assessment.
• To address the need for reliable and efficient objective quality evaluation of distorted 3D point clouds.
• To improve the accuracy of quality assessment for 3D models used in various applications.

Main Methods:

• Extraction of geometric and perceptual attributes from distorted 3D point clouds.
• Representation of attributes as 1D vectors for feature extraction.
• Application of transfer learning with a 1D convolutional neural network (1D CNN) adapted from 2D CNNs.
• Quality score prediction using regression with fully connected layers.

Main Results:

• The proposed NR method demonstrates superior performance in 3D point cloud quality assessment.
• The approach shows enhanced correlation with average opinion scores across multiple datasets.
• Evaluated on SJTU_PCQA, WPC, and ICIP2020 databases, achieving state-of-the-art results.

Conclusions:

• The deep learning-based NR method provides an effective solution for 3D point cloud quality assessment.
• The methodology offers a reliable and efficient way to evaluate distortions in 3D models.
• This work contributes to the advancement of objective quality evaluation for 3D data.