Asymmetric dual-prior multitask learning for no-reference point cloud quality assessment

Point clouds have become an important representation for three-dimensional vision and computer graphics applications. However, they are susceptible to various distortions during acquisition, compression, storage, and transmission, making accurate point-cloud quality assessment essential. Although existing no-reference point-cloud quality assessment methods have explored multitask learning, most of them exploit only one type of auxiliary information, either distortion attributes or perceptual quality levels. The complementary relationships between these two heterogeneous priors and continuous quality prediction have not been fully exploited. To address this limitation, we propose an asymmetric dual-prior multitask learning method for projection-based no-reference point-cloud quality assessment. Leveraging an established six-view multiattribute projection pipeline that outputs texture, depth, and occupancy maps, our approach integrates customized modules and loss designs for multitask cooperation. We employ a scale-token cross-level interaction (STCI) module, which encodes hierarchical features into scale tokens to capture dependencies among adjacent and nonadjacent levels. For bridging auxiliary tasks and regression targets, two asymmetric cross-task constraints are developed, including a distortion-type-aware quality ranking constraint for unordered distortion labels and a regression-level consistency constraint for ordered quality-level labels. We also introduce a primary-task-oriented progressive weighting strategy that gradually shifts the optimization focus away from auxiliary supervision toward continuous quality regression. Experimental results on the SJTU-PCQA and WPC databases demonstrate that the proposed method achieves competitive quality prediction performance. Ablations show consistent gains from STCI, the two cross-task constraints, and progressive weighting.