ORNet: No-reference Point Cloud Quality Assessment in an Ordinal Regression Way

No-reference Point Cloud Quality Assessment (NR-PCQA) is pivotal for automatically evaluating the perceptual visual quality of point clouds, with the goal of accurately aligning model predictions with ground-truth human subjective ratings. However, a common oversight of traditional NR-PCQA models is their tendency to neglect the inherent ordering relationships that exist between different quality ratings. This frequently results in a gap between the algorithm's predictions and the nuanced reality of human perception. Drawing inspiration from the process of subjective human rating of visual quality, which evolves from uncertainty to certainty, we propose reformulating NR-PCQA as an ordinal regression problem. Our approach involves two key steps. First, utilizing an ordinal loss function for NR-PCQA learning to formulate ordinal regression as binary classification sub-problem; Second, transforming predicted probabilities into quality scores through continuous-based inference which can mitigates discretization errors and yields more realistic quality scores compared to traditional inference strategy. Extensive experiments on the SJTU-PCQA, WPC, and WPC2.0 datasets confirm the superior performance of our proposed ORNet in comparison with state-of-the-art methods. The project can be found in https://github.com/YovenLi/ORNet.