Denoising Diffusion Probabilistic Model for Point Cloud Compression at Low Bit-Rates

TL;DR

The paper tackles very low bitrate point cloud compression by leveraging a diffusion-based autoencoder (DDPM-PCC) where a PointNet encoder produces a latent $\mathbf{z}$ that is quantized with a learnable codebook into $\hat{\mathbf{z}}$, and this quantized latent conditions the reverse diffusion decoder to reconstruct the geometry. The latent is divided into $C$ chunks and each chunk is quantized with a codebook of size $N=128$, yielding a rate of $C\log_2 N$ bits; the model is trained with a joint objective $\mathcal{L}=\mathcal{L}_{\text{Diff}}+\mathcal{L}_{\text{VQ}}$ using $T=200$ steps and a linear $\beta$ schedule. Experimental results on ShapeNet and ModelNet40 show superior rate-distortion performance at low bitrates compared to G-PCC, Draco, and prior learning-based PCC methods, with ablations confirming the benefits of VQ over fully factorized latent modeling and the encoder choices. The approach demonstrates a practical diffusion-based pathway to high-fidelity point cloud reconstruction under tight bandwidth constraints, and the authors provide public code to encourage reproducibility.

Abstract

Efficient compression of low-bit-rate point clouds is critical for bandwidth-constrained applications. However, existing techniques mainly focus on high-fidelity reconstruction, requiring many bits for compression. This paper proposes a "Denoising Diffusion Probabilistic Model" (DDPM) architecture for point cloud compression (DDPM-PCC) at low bit-rates. A PointNet encoder produces the condition vector for the generation, which is then quantized via a learnable vector quantizer. This configuration allows to achieve a low bitrates while preserving quality. Experiments on ShapeNet and ModelNet40 show improved rate-distortion at low rates compared to standardized and state-of-the-art approaches. We publicly released the code at https://github.com/EIDOSLAB/DDPM-PCC.

Figures (5)

• Figure 1: Low bit-rates PCC and compression artifacts.
• Figure 2: Overview of the proposed method. Here for simplicity, $\mathbf{x}^{(t-1)}$ represents a single point, while the diffusion process shows the final reconstructed point cloud (conditioned by the quantized latent representation $\mathbf{\hat{z}}$). Here $\mathbf{z}$ is split into $C = 4$ chunks, making $\mathbf{\hat{z}}$ representable with only $4$ codebook indices.
• Figure 3: Quantitative results on ShapeNet (first row) and ModelNet40 (second row). The y-axes are plotted in logarithmic scale. Our method achieves the best geometry reconstruction at a very low bit rate according to all quality metrics.
• Figure 4: Qualitative results on ModelNet (first two rows) and ShapeNet (last two rows). We compare the reconstruction results of different compression methods providing the bpp and the PSNR values for each analyzed sample.
• Figure 5: Comparison between a learnable vector quantize approach (VQ) and a fully factorize-based strategy (FF) to encode the latent representation. The dashed line represents the lower bound in Chamfer Distance: the generative model. Different values of $C$ allow for different RD trade-offs