UPDA: Unsupervised Progressive Domain Adaptation for No-Reference Point Cloud Quality Assessment

TL;DR

UPDA tackles domain shift in NR-PCQA by introducing a two-stage unsupervised progressive domain adaptation framework. It first performs discrepancy-aware coarse-grained alignment (DACA) to transfer relative quality rankings via ranking prediction and discrepancy-weighted MMD, then applies perception fusion fine-grained alignment (PFFA) with symmetric feature fusion and a conditional discriminator to transfer absolute quality scoring. The approach preserves the original model architecture at inference and improves cross-domain generalization across multiple NR-PCQA backbones on SJTU-PCQA, WPC, and WPC 2.0, as shown by extensive cross-distortion and cross-dataset experiments and thorough ablations. The method achieves practical applicability with modest training overhead and provides code for reproducibility, suggesting potential extensions to source-free and test-time adaptation. Overall, UPDA represents a principled, two-stage strategy that leverages ranking signals and feature fusion to robustly generalize NR-PCQA across diverse domains.

Abstract

While no-reference point cloud quality assessment (NR-PCQA) approaches have achieved significant progress over the past decade, their performance often degrades substantially when a distribution gap exists between the training (source domain) and testing (target domain) data. However, to date, limited attention has been paid to transferring NR-PCQA models across domains. To address this challenge, we propose the first unsupervised progressive domain adaptation (UPDA) framework for NR-PCQA, which introduces a two-stage coarse-to-fine alignment paradigm to address domain shifts. At the coarse-grained stage, a discrepancy-aware coarse-grained alignment method is designed to capture relative quality relationships between cross-domain samples through a novel quality-discrepancy-aware hybrid loss, circumventing the challenges of direct absolute feature alignment. At the fine-grained stage, a perception fusion fine-grained alignment approach with symmetric feature fusion is developed to identify domain-invariant features, while a conditional discriminator selectively enhances the transfer of quality-relevant features. Extensive experiments demonstrate that the proposed UPDA effectively enhances the performance of NR-PCQA methods in cross-domain scenarios, validating its practical applicability. The code is available at https://github.com/yokeno1/UPDA-main.

Figures (4)

• Figure 1: Performance degradation problem of typical NR-PCQA methods (MM-PCQA zhang2023mm, GMS-3DQA zhang2024gms, and 3DTA zhu20243dta) on cross-domain datasets. The source domain (e.g., SJTU-PCQA yang21sjtu) and target domain (e.g., WPC su2019perceptual) exhibit significant disparities in both content characteristics and distortion types.
• Figure 2: The overall structure of the proposed network. Firstly, in the discrepancy-aware coarse-grained alignment (DACA) stage, we train the feature extractor $G$ by introducing quality difference information. Then, in the perception fusion fine-grained alignment stage, we implement a hierarchical refinement process to progressively bridge domain gaps while preserving quality-sensitive features by symmetric feature fusion and conditional discriminator.
• Figure 3: t-SNE visualization of the feature space (red: source domain feature, blue: target domain feature).
• Figure 4: Parameter sensitivity analysis of $\nu$ and $\mu$. The baseline model is GMS-3DQA zhang2024gms.