ProDAT: Progressive Density-Aware Tail-Drop for Point Cloud Coding

TL;DR

This work tackles efficient progressive coding for large-scale 3D point clouds by introducing ProDAT, a density-aware tail-drop framework that enables progressive decoding with a single trained model. By ranking channels with density-guided importance and jointly dropping latent features and coordinates, ProDAT achieves rate-distortion efficiency across bitrates, with a progressive ratio $\alpha \in [0,1]$ controlling reconstruction fidelity. The approach combines a density-based drop rule $\rho = \rho_{\max} - (\rho_{\max}-\rho_{\min}) \delta$, channel-importance $\mathcal{I}_c = \beta \cdot \mathrm{norm}(\mathrm{Var}_c) + (1-\beta) \cdot \mathrm{norm}(\mathrm{Grad}_c)$ with $\beta=0.6$, and an integrated loss that balances Chamfer Distance, density preservation, coordinate regularization, and bitrate. Experiments on SemanticKITTI and ShapeNet show substantial BD-rate gains over state-of-the-art methods (e.g., over $28.6\%$ on SemanticKITTI for PSNR-D2 and over $18.15\%$ on ShapeNet), validating both progressive decoding capability and coding efficiency for real-time 3D transmission and rendering.

Abstract

Three-dimensional (3D) point clouds are becoming increasingly vital in applications such as autonomous driving, augmented reality, and immersive communication, demanding real-time processing and low latency. However, their large data volumes and bandwidth constraints hinder the deployment of high-quality services in resource-limited environments. Progres- sive coding, which allows for decoding at varying levels of detail, provides an alternative by allowing initial partial decoding with subsequent refinement. Although recent learning-based point cloud geometry coding methods have achieved notable success, their fixed latent representation does not support progressive decoding. To bridge this gap, we propose ProDAT, a novel density-aware tail-drop mechanism for progressive point cloud coding. By leveraging density information as a guidance signal, latent features and coordinates are decoded adaptively based on their significance, therefore achieving progressive decoding at multiple bitrates using one single model. Experimental results on benchmark datasets show that the proposed ProDAT not only enables progressive coding but also achieves superior coding efficiency compared to state-of-the-art learning-based coding techniques, with over 28.6% BD-rate improvement for PSNR- D2 on SemanticKITTI and over 18.15% for ShapeNet

Figures (9)

• Figure 1: Comparisons between Point Cloud Coding (PCC) and Progressive Point Cloud Coding (PPCC). PPCC generates a progressive bitstream post-encoding, enabling customized decoding for variable coding ratios $a$ and reconstruction qualities.
• Figure 2: The architecture of the proposed ProDAT. The progressive coding model of ProDAT leverages an end-to-end autoencoder which consists of an Encoder $E$, a Density-aware Tail-drop Operator, an Entropy Bottleneck, and a Decoder $D$. The total loss is defined as a trade-off among density loss, Chamfer Distance (CD) loss, and bitrate. See Sect. \ref{['ProDAT_Framework']} for further details.
• Figure 3: Quantitative results of non-progressive ProDAT on SemanticKITTI and ShapeNet: BPP vs. CD (a, c) and BPP vs. PSNR-D2 (b, d). Models are trained using progressive coding. Results compare two tail-drop strategies: Combined Drop (red lines) and Feature-Only Drop (orange lines), with the model trained using progressive coding. The differences between these strategies will be described in terms of the Ablation Study below.
• Figure 4: Quantitative comparison of non-progressive ProDAT with JPEG-VM guarda2022ipleiria on SemanticKITTI in terms of PSNR-D1 and PSNR-D2 for each point cloud model.
• Figure 5: Progressive performance of ProDAT (PSNR-D2 vs BPP) on SemanticKITTI and ShapeNet across different $\lambda$ values with simultaneous latent feature and coordinate drop.