[Paper Review] Progressive Knowledge Transfer Based on Human Visual Perception Mechanism for Perceptual Quality Assessment of Point Clouds

Qi Liu, Yiyun Liu+3 more|arXiv (Cornell University)|Nov 30, 2022

Image and Video Quality Assessment4 citations

TL;DR

This paper proposes PKT-PCQA, a no-reference perceptual quality assessment model for point clouds that leverages a progressive knowledge transfer framework inspired by the human visual system (HVS). By integrating spatial and channel attention mechanisms with graph spectral features and hierarchical feature fusion, PKT-PCQA achieves state-of-the-art performance on three benchmark datasets, outperforming existing no-reference and reduced-reference methods while matching full-reference baselines in correlation with subjective scores.

ABSTRACT

With the wide applications of colored point cloud in many fields, point cloud perceptual quality assessment plays a vital role in the visual communication systems owing to the existence of quality degradations introduced in various stages. However, the existing point cloud quality assessments ignore the mechanism of human visual system (HVS) which has an important impact on the accuracy of the perceptual quality assessment. In this paper, a progressive knowledge transfer based on human visual perception mechanism for perceptual quality assessment of point clouds (PKT-PCQA) is proposed. The PKT-PCQA merges local features from neighboring regions and global features extracted from graph spectrum. Taking into account the HVS properties, the spatial and channel attention mechanism is also considered in PKT-PCQA. Besides, inspired by the hierarchical perception system of human brains, PKT-PCQA adopts a progressive knowledge transfer to convert the coarse-grained quality classification knowledge to the fine-grained quality prediction task. Experiments on three large and independent point cloud assessment datasets show that the proposed no reference PKT-PCQA network achieves better of equivalent performance comparing with the state-of-the-art full reference quality assessment methods, outperforming the existed no reference quality assessment network.

Motivation & Objective

• To address the gap in existing point cloud quality assessment methods that fail to model human visual system (HVS) properties, leading to inaccurate perceptual predictions.
• To develop a no-reference quality assessment framework that does not require access to the original (reference) point cloud, enabling broader real-world deployment.
• To improve the alignment between objective model predictions and human subjective opinion scores by incorporating HVS-inspired mechanisms such as selective attention and hierarchical processing.
• To design a progressive knowledge transfer strategy that transfers coarse-grained quality classification knowledge to fine-grained quality prediction, enhancing model generalization and accuracy.

Proposed method

• The model employs a graph spectral-based feature extraction mechanism to capture global geometric structure and key cluster information from point clouds, reducing computational complexity.
• Local and global features are fused through a hierarchical architecture that mimics the human visual system’s multi-scale receptive fields, enhancing representation learning.
• A dual attention mechanism—spatial and channel attention—is integrated to dynamically weight important features, emulating human selective visual attention for improved quality prediction.
• A progressive knowledge transfer strategy is implemented, transferring knowledge from a coarse quality classification task to a fine-grained quality regression task, improving model generalization.
• The network is trained end-to-end using logistic regression for score scaling to align objective scores with subjective MOS (Mean Opinion Score) ranges.
• The framework is evaluated on three independent datasets: WPC, SJTU-PCQA, and M-PCCD, using standard metrics including PLCC and SROCC.

Experimental results

Research questions

• RQ1 Can a no-reference point cloud quality assessment model achieve performance comparable to full-reference methods by modeling human visual perception?
• RQ2 How does incorporating spatial and channel attention mechanisms improve the alignment between objective quality scores and human subjective opinions?
• RQ3 To what extent does progressive knowledge transfer from coarse classification to fine-grained prediction enhance model accuracy in perceptual quality assessment?
• RQ4 Can a graph spectral-based feature extraction method effectively capture both local and global geometric features in irregular point cloud data?
• RQ5 How does the proposed model perform across diverse datasets with varying distortion levels and point cloud characteristics?

Key findings

• On the WPC dataset, PKT-PCQA achieved a PLCC of 0.56 and SROCC of 0.56, the highest among all no-reference and reduced-reference methods, and only slightly below the best full-reference method IW-SSIMp.
• On the SJTU-PCQA dataset, PKT-PCQA achieved a PLCC of 0.91 and SROCC of 0.91, significantly outperforming other no-reference models and even surpassing the reduced-reference PCM_RR metric.
• On the M-PCCD dataset, PKT-PCQA achieved a SROCC of 0.91 and PLCC of 0.93, demonstrating strong generalization across diverse point cloud types and distortions.
• The model’s performance was consistently superior to the state-of-the-art no-reference method PQA-NET, which achieved only PLCC 0.71 and SROCC 0.77 on the WPC dataset.
• The proposed method achieved near-state-of-the-art performance—within 0.02–0.05 of the best full-reference methods—despite not requiring any reference data, highlighting its robustness and practicality.
• The integration of HVS-inspired attention mechanisms and progressive knowledge transfer significantly improved prediction accuracy, as evidenced by higher correlation with subjective MOS across all datasets.