Towards Perceptual Fidelity: KDE-Peak-Based JND Modeling for Point Cloud Attribute Compression

Just Noticeable Difference (JND) plays a critical role in visual media compression oriented towards human visual perception. JND models help in improving compression efficiency while ensuring the preservation of visual quality. While existing methods for JND modeling for images/videos (e.g., kmeans clustering, Gaussian Mixture Models) demonstrate their effectiveness in general scenarios, they encounter the following chanlleges when applied to perceptual lossless point cloud compression focused on first JND localization: reliance on distribution assumptions, cross-level data aliasing from global clustering, and poor adaptability to small-sample or non-normal datasets. To address these challenges, we propose a Kernel Density Estimation (KDE) peak-based method for JND modeling that eliminates distributional priors through non-parametric density fitting. The major contributions of the proposed work include: (1) frequency-weighted kernel smoothing to amplify perceptual consensus among subjects, (2) hierarchical density estimation (1st_JND-nth_JND) to suppress cross-level interference, and (3) a dual-objective optimization strategy balancing bitstream loss against perceptual error for optimal quantization parameter (QP) selection. Experimental results demonstrate that the proposed method accurately identifies JND levels, providing an efficient and robust JND modeling method for perceptual lossless point cloud attribute compression.