Just Noticeable Difference Measurement for Point Cloud Compression: A Benchmark Dataset and Prediction Network

As point clouds have become an increasingly important representation for immersive visualization of real-world environments, perceptual quality measurement has become a key issue in point cloud sensing and compression systems. Just Noticeable Distortion (JND) provides an effective perceptual criterion for characterizing the minimum distortion that can be perceived by the human visual system. However, progress in point cloud JND modeling has been largely hindered by the lack of large-scale, high-quality subjective datasets. To address this issue, we present PKU-JND, a new large-scale point cloud JND dataset constructed through carefully designed subjective experiments. PKU-JND contains 230 pristine point cloud models with diverse spatial distributions, from which 5,750 geometry-distorted and 7,130 attribute-distorted samples are generated. Each JND threshold is annotated by 20–31 participants over a five-month period, and rigorous post-processing procedures are applied to ensure annotation reliability. Based on the proposed dataset, we further develop a novel JND prediction baseline network, termed JND-Net, which models perceptual distortion thresholds for point clouds. Experimental results show that JND-Net achieves PLCC/SRCC/KRCC/RMSE scores of 0.549/0.459/0.326/4.398 for geometry JND prediction and 0.465/0.439/0.295/4.202 for attribute JND prediction. Furthermore, when the predicted JND points are integrated into point cloud compression, JND-Net consistently achieves encoding results closest to the annotated perceptual thresholds and obtains lower bitrate at comparable perceptual quality levels, demonstrating improved rate-distortion efficiency for perceptually guided point cloud coding. The dataset is released at https://github.com/Heroism502/PKU-JND.