UGAE: Unified Geometry and Attribute Enhancement for G-PCC Compressed Point Clouds

TL;DR

UGAE addresses the coupled distortions from lossy G-PCC point-cloud compression by introducing a three-part framework: PoGE to refine geometry, PAE to recolor attributes using the enhanced geometry at the encoder, and PoAE to refine attributes at the decoder with a high-frequency–focused W‑MSE loss. The approach leverages a Transformer-based sparse U‑Net for geometry, DA‑KNN recoloring for high-frequency preservation, and a residual attribute predictor guided by frequency-aware loss, achieving large improvements over G-PCC TMC13v29 on benchmarks such as $8iVFB$, $Owlii$, and $MVUB$ (e.g., $D1$ BD‑PSNR gains up to $9.98$ dB and BD‑BR reductions up to $-90.98\%$ for geometry; $Y$ BD‑PSNR gains of $3.67$ dB for attributes). By enforcing deterministic geometry outputs through CPU-based TSConv and integrating geometry enhancement into attribute reconstruction, UGAE substantially improves both objective metrics and perceptual quality. These results demonstrate a practical pathway to higher-fidelity compressed point clouds, with potential impact on streaming, storage, and downstream 3D processing tasks.

Abstract

Lossy compression of point clouds reduces storage and transmission costs; however, it inevitably leads to irreversible distortion in geometry structure and attribute information. To address these issues, we propose a unified geometry and attribute enhancement (UGAE) framework, which consists of three core components: post-geometry enhancement (PoGE), pre-attribute enhancement (PAE), and post-attribute enhancement (PoAE). In PoGE, a Transformer-based sparse convolutional U-Net is used to reconstruct the geometry structure with high precision by predicting voxel occupancy probabilities. Building on the refined geometry structure, PAE introduces an innovative enhanced geometry-guided recoloring strategy, which uses a detail-aware K-Nearest Neighbors (DA-KNN) method to achieve accurate recoloring and effectively preserve high-frequency details before attribute compression. Finally, at the decoder side, PoAE uses an attribute residual prediction network with a weighted mean squared error (W-MSE) loss to enhance the quality of high-frequency regions while maintaining the fidelity of low-frequency regions. UGAE significantly outperformed existing methods on three benchmark datasets: 8iVFB, Owlii, and MVUB. Compared to the latest G-PCC test model (TMC13v29), UGAE achieved an average BD-PSNR gain of 9.98 dB and 90.98% BD-bitrate savings for geometry under the D1 metric, as well as a 3.67 dB BD-PSNR improvement with 56.88% BD-bitrate savings for attributes on the Y component. Additionally, it improved perceptual quality significantly.

Figures (8)

• Figure 1: UGAE pipeline. At the encoder side, PoGE enhances the lossy geometry, and PAE recolors the enhanced geometry using the original attribute information. At the decoder side, PoGE reconstructs the same enhanced geometry as in the encoder side to assist in attribute decoding, and PoAE focuses on high-frequency regions to produce the final enhanced point cloud.
• Figure 2: PoGE architecture. The initialization converts the unstructured point cloud into a structured format, and PT blocks extract multi-scale features. The enhanced geometry is obtained from the geometry enhancement head (Geo-Enh Head) through probability-based sorting and selection.
• Figure 3: PoAE architecture. PT blocks extract multi-scale features, and the attribute enhancement head (Att-Enh Head) predicts color residuals, which are added to the original attributes to obtain the enhanced attributes.
• Figure 4: Illustration of DA-KNN recoloring for $k=8$. The left side shows a query point (shown in blue) in the enhanced geometry along with its neighboring points (shown in yellow and orange); the right side illustrates the spatial distribution of neighbors in 3D space. DA-KNN first finds the $k$ nearest neighbours (shown in yellow and orange) of the query point (shown in blue). Then, from these $k$ neighbours, the algorithm selects the closest ones that lie at the same distance from the query point. In this example, $k_t=3$ points (shown in orange) are selected.
• Figure 5: (a) High-frequency (top 50%) regions in the original point cloud attributes. (b) High-loss (top 50%) regions in the decoded point cloud attributes. (c) Overlap between (a) and (b): yellow indicates overlapping areas, while blue and light red denote non-overlapping regions; the overlap ratio is 75.09%. (d) Overlap visualization after PoAE, with an overlap ratio of 53.08%.