PaperlayerView exact revision on arXiv

Source-linked AI summary

Point Cloud Transmission With Patch-Based Rate Adaptation

Shipeng Liu, Garima Bazard, Khac-Hoang Ngo, Per-Erik Forssén

arXiv:2610.29479v1cs.ITeess.SPmath.IT

TL;DR

Wireless point-cloud transmission faces large data volumes, irregular geometry, and channel variation, while patch-based JSCC methods may assign equal rates despite differences and dependencies among patches. GeoContext-JSCC uses anchor-conditioned entropy estimates to allocate variable symbol lengths to patches. On ModelNet40, it improves rate–distortion performance over the considered baselines across transmission rates and channel conditions.

  • Problem

    Equal-rate patch-based JSCC does not account for local geometry, and correlated, overlapping patches complicate entropy estimation for efficient channel allocation.

    Source: §I · p. 1

  • Method

    GeoContext-JSCC uses anchor-patch context to estimate conditional entropy and assign variable-length transmission to patch latents, with analog symbols and digitally transmitted centroids and rate indicators.

    Source: §I · p. 1, §III · p. 2, 4) Conditional Entropy and Patch-Based… · p. 4

  • Results

    ModelNet40 experiments show consistent rate–distortion improvements over G-PCC, DPCC, existing point-cloud JSCC methods, and the fixed-rate variant across transmission rates and channel conditions.

    Source: §I · p. 1

  • Takeaways & Limitations

    Ablations show that both variable-rate allocation and anchor-conditioned entropy modeling contribute to reconstruction performance, while the framework maintains graceful degradation as channel quality decreases.

    Source: §I · p. 1, §V · p. 6

Abstract

from arXiv · show

We propose a joint source-channel coding (JSCC) framework for efficient wireless point cloud transmission. In the proposed scheme, the point cloud is first divided into local patches and transformed into compact latent representations, which are then mapped to symbols for transmission over a noisy channel. Our model incorporates an entropy model that exploits inter-patch spatial dependencies to estimate patch-level coding costs and allocate different numbers of symbols accordingly. Unlike existing patch-based point cloud JSCC methods that use the same transmission rate between patches, we adapt the channel usage of each patch latent representation. To exploit inter-patch dependencies without patch-by-patch autoregressive processing, we select a group of anchor patches whose features provide contextual information for entropy estimation of non-anchor ones. The framework is trained using a reconstruction loss and an entropy-based rate surrogate. Experiments on ModelNet40 demonstrate overall rate-distortion improvements over existing baselines across channel conditions.

I. INTRODUCTION

Wireless point-cloud transmission must handle large data volumes, irregular geometry, and channel variation; GeoContext-JSCC adapts patch transmission using spatial context. On ModelNet40, it improves rate–distortion performance over the considered baselines across rates and channel conditions.

I. INTRODUCTION

  • Point clouds’ large data volume, irregular sparse structure, and sensitivity to geometric distortion complicate real-time wireless transmission.Conventional separate source-channel coding can exhibit abrupt quality degradation below its designed channel condition and saturation once decoding is reliable.Source: §I · p. 1
  • Existing patch-based JSCC assigns equal-length symbol sequences to patches, while correlated neighboring patches make accurate entropy estimation difficult for adaptive allocation.Overlapping neighborhoods can contain partially redundant geometry, and ignoring these dependencies can make channel allocation inefficient.Source: §I · p. 1
  • ModelNet40 experiments show consistent rate–distortion improvements over G-PCC, DPCC, existing point-cloud JSCC methods, and the fixed-rate variant across transmission rates and channel conditions.Ablations attribute reconstruction contributions to entropy-guided channel allocation and inter-patch contextual coding; the framework also maintains graceful degradation as channel quality decreases, unlike G-PCC- and DPCC-based methods.Source: §I · p. 1

II. SYSTEM MODEL AND PROBLEM FORMULATION

  • The system maps an input point cloud to channel symbols and optimizes encoder and decoder for a rate–distortion objective under an average-power constraint.The channel model uses AWGN after equalization, and reconstruction distortion is measured with Chamfer distance.Source: §II · p. 2

III. PROPOSED SCHEME

GeoContext-JSCC uses estimated conditional entropy to set variable transmission lengths for patch latents. It combines analog transmission of latent symbols with digital transmission of patch centroids and rate indicators.

III. PROPOSED SCHEME

  • GeoContext-JSCC maps patch latents to variable-length symbol sequences, with lengths determined by estimated conditional entropy.The symbols are transmitted in an analog manner, while patch centroids and rate-allocation indicators use standard digital communication.Source: §III · p. 2
  • Anchor-patch features provide contextual information for non-anchor patches, keeping the inter-patch entropy model tractable without patch-by-patch processing.At the receiver, decoded local patch geometries are translated by recovered centroids and assembled into the reconstructed point cloud.Source: §III · p. 2
  • Each local patch is represented by a centroid and a latent vector, and its latent symbols are transmitted separately from the digital centroid and rate indicator.Source: §III · p. 2

A. Patch Construction and Representation

The point cloud is divided into local patches using farthest point sampling and nearest-neighbor grouping. Each centered patch is encoded as a latent representation paired with its centroid.

A. Patch Construction and Representation

B. JSCC Transmission of Patch Latents

Anchor patches provide shared spatial context for non-anchor patches. The Cross-Point Transformer uses nearby patch features to form context-enhanced representations in the framework’s modules.

1) Anchor-Conditioned Context:

  • The first SA patches selected by farthest point sampling serve as anchors, while the remaining S − SA patches are non-anchors conditioned on anchor latents.Anchor patches are processed independently in parallel before non-anchor patches; the figure also indicates reconstructed anchor features support non-anchor decoding.Source: 1) Anchor-Conditioned Context · p. 2, Figure 1 · p. 3

2) Cross-Point Transformer (CPT):

  • For each query patch, the Cross-Point Transformer selects nearby key/value patches by centroid distance and uses learned projections and relative-position embeddings.Source: 2) Cross-Point Transformer (CPT) · p. 3
  • Position-aware attention aggregates neighboring projected features into context-enhanced query features, and the CPT is used in the spatial context encoder and decoder and the conditional entropy model.Source: 2) Cross-Point Transformer (CPT) · p. 3

3) Non-Anchor Spatial Context Encoder and Decoder:

The non-anchor context modules use anchor patches to form residual representations for transmission and combine recovered residuals with reconstructed anchors at the receiver. The resulting latents are passed to the patch feature decoder for local geometry reconstruction.

3) Non-Anchor Spatial Context Encoder and Decoder:

4) Conditional Entropy and Patch-Based Rate Adaptation:

The section models patch latents with Gaussian distributions and uses nearby anchor context to estimate non-anchor distributions. Estimated patch coding costs determine each patch’s allocated channel length.

4) Conditional Entropy and Patch-Based Rate Adaptation:

  • Anchor latents use a learned Gaussian distribution, while non-anchor latents use a learned Gaussian distribution conditioned on neighboring anchor patches.Source: 4) Conditional Entropy and Patch-Based… · p. 3
  • For each non-anchor patch, the nearest anchor patches by centroid distance provide context used to predict its Gaussian parameters.An MLP predicts the mean and standard deviation from the anchor-conditioned feature.Source: 4) Conditional Entropy and Patch-Based… · p. 4
  • The estimated coding cost sets a continuous rate surrogate, and the allocated patch length is the closest value in the available rate set.The surrogate is αR_i; the selected length is drawn from K.Source: 4) Conditional Entropy and Patch-Based… · p. 4

5) Rate-Adaptive Latent Transmission:

Rate-adaptive transmission maps patch representations to channel symbols using selected rate levels, with embeddings conditioning the encoder and decoder inputs.

5) Rate-Adaptive Latent Transmission:

6) Training JSCC:

Training replaces the CBR term with a differentiable entropy-based rate surrogate and proceeds in two stages: noise-free autoencoder pretraining, then joint training of the remaining modules with the feature encoder and decoder frozen.

6) Training JSCC:

  • The training objective replaces CBR with the differentiable surrogate ∑_i αR_i.Source: 6) Training JSCC · p. 4
  • Stage 1 pretrains the feature encoder and decoder as a noise-free autoencoder using patch latent representations.Source: 6) Training JSCC · p. 4
  • In Stage 2, the feature encoder and decoder are frozen while the remaining modules are jointly trained.Source: 6) Training JSCC · p. 4

C. Digital Transmission of Centroids and Rate Allocation

The digital transmission accounting covers patch centroids and rate-allocation indicators, alongside the expected channel-use cost under retransmission. The combined analog–digital CBR varies with SNR.

C. Digital Transmission of Centroids and Rate Allocation

  • Centroid coordinates and rate-allocation indicators are sorted in the same order within the anchor and non-anchor groups, then sent over digital links.Representing each centroid coordinate with b bits requires 3Sb bits in total.Source: C. Digital Transmission of Centroids… · p. 4
  • The total-payload accounting includes signaling for rate-allocation count vectors for both anchor and non-anchor patches.The passage introduces the count-vector counts for the two groups and then describes the total payload.Source: C. Digital Transmission of Centroids… · p. 4
  • With independent retransmissions, perfect error detection, and negligible feedback overhead, the expected channel uses are n/(1 − ϵ(B, n, SNR)), minimized over n.The block error probability ϵ(B, n, SNR) is for transmitting B bits over a complex-valued AWGN channel with blocklength n.Source: C. Digital Transmission of Centroids… · p. 4
  • The combined analog–digital CBR varies with SNR.The passage introduces a CBR computation after combining analog and digital transmission.Source: C. Digital Transmission of Centroids… · p. 4

IV. NUMERICAL EXPERIMENTS

On ModelNet40, GeoContext-JSCC outperforms the considered methods across tested CBRs and SNRs, with larger gains at lower CBR. The experiments also compare fixed-rate allocation and evaluate the effects of that choice.

A. Experimental Setup

  • Each ModelNet40 point cloud contains 8192 points, and the patch latent dimension is 64.The supported channel-use levels are K = {2, 4, 6, 8, 12, 16, 20, 24, 28, 32}; the first half of the patches are designated as anchors.Source: A. Experimental Setup · pp. 4–5
  • Both training stages last 40 epochs, with Stage 2 sampling channel SNR from 0 to 10 dB; communication cost is reported via CBR and reconstruction quality via D1 and D2 PSNR.The rate–distortion trade-off coefficient is λ = 10−4.Source: A. Experimental Setup · p. 5

B. Rate–Distortion Performance

  • The comparison includes G-PCC, SEPT, DPCC, and Xie et al., with CBR varied through patch count, patch rate, or compression rate.G-PCC and DPCC bitstreams use rate-1/2 polar coding and QPSK; SEPT is plotted only at CBR = 0.017 because higher-CBR configurations did not improve D1 or D2 PSNR in the reproduction.Source: B. Rate–Distortion Performance · p. 5
  • At SNR = 10 dB, GeoContext-JSCC gains 0.55–1.65 dB in D1 PSNR and 1.05–2.67 dB in D2 PSNR over Xie et al..It outperforms the other considered methods across the tested CBR range and SNR ∈{5, 10} dB; gains are more pronounced at lower CBRs.Source: B. Rate–Distortion Performance · p. 5

C. Ablation Studies; 1) Effect of Variable-Rate Allocation:

  • The ablation comparison uses S = 64 patches, or K = 256 points per patch, with the same latent dimension, training schedule, channel settings, and reconstruction backbone.The fixed-rate variant assigns every patch a sequence length of k = 3 complex values, yielding an average CBR comparable to the rate-allocation model.Source: C. Ablation Studies · p. 5, 1) Effect of Variable-Rate Allocation · p. 5

2) Effect of Spatial Context:

Ablations show that variable-rate allocation and anchor-based spatial context each contribute to reconstruction quality. The visual comparison and conclusion describe the framework’s performance and scope on ModelNet40.

2) Effect of Spatial Context:

  • The full model with rate allocation and spatial context outperforms variants without either component in D1 and D2 PSNR across tested SNRs.The reported CBR decreases from approximately 0.142 at 0 dB to 0.046 at 10 dB; the authors describe the two components as providing complementary gains.Source: 2) Effect of Spatial Context · p. 5, Figure 4 · p. 6

D. Visual Comparison

  • At CBR ≈0.017 and SNR = 10 dB, GeoContext-JSCC better preserves object shapes and distributes reconstructed points more uniformly than Xie et al. and SEPT.Fig. 5 shows ground truth, GeoContext-JSCC, Xie et al., and SEPT from left to right.Source: D. Visual Comparison · p. 6, Figure 5 · p. 6

V. CONCLUSION

  • GeoContext-JSCC uses anchor-derived context to estimate patch-latent conditional entropy and determine per-patch transmission lengths.The conclusion describes the framework as contextual variable-rate JSCC for wireless point-cloud transmission.Source: §V · p. 6
  • On ModelNet40 over AWGN channels, the framework improves rate–distortion performance over the considered conventional digital and learning-based JSCC baselines.The conclusion also says ablations confirm contributions from variable-rate allocation and anchor-conditioned entropy modeling.Source: §V · p. 6
Loading 2610.29479v1…