DEEP LOSSLESS POINT CLOUD ATTRIBUTE COMPRESSION VIA EED PREDICTION

Lossless point cloud attribute compression remains challenging due to irregular geometry, varying densities, and the limited prediction strategies in existing learning-based approaches. We propose DeepEED, a framework that integrates edge-enhancing anisotropic diffusion (EED) with adaptive deep entropy coding. Hybrid hierarchical partitioning separates base and inference points, followed by LoD partitioning and uniform sampling. Base layers are compressed using point-wise prediction and run-length coding, while inference points are iteratively predicted via EED. Residuals are compressed layer-wise using a local adaptive entropy model, with reconstructed attributes guiding subsequent prediction and coding. Experiments on diverse datasets show that DeepEED achieves 12.6%-15.2% bitrate reduction over MPEG G-PCC and outperforms learned methods, while enabling faster decoding and stronger generalization. © 2026 IEEE