V2X-DSC: Multi-Agent Collaborative Perception with Distributed Source Coding Guided Communication

TL;DR

The paper tackles the challenge of bandwidth-limited multi-agent collaborative perception by reframing it through distributed source coding. It introduces V2X-DSC, a DSC-guided Conditional Codec that encodes only the innovation in a collaborator's BEV feature relative to the receiver's local context, leveraging side information at the decoder. The approach combines sender-side pruning, discrete bottleneck coding, entropy coding, and a receiver-side side-information network to achieve fusion-ready reconstructions under kilobyte-level per-link bandwidth, while maintaining or improving detection accuracy across DAIR-V2X, OPV2V, and V2X-Real. Extensive experiments show state-of-the-art accuracy–bandwidth trade-offs, plug-and-play compatibility with multiple fusion backbones, and robustness to pose noise and communication delays. This work demonstrates the practical value of information-theoretic principles for designing communication-efficient perception systems in connected autonomous environments.

Abstract

Collaborative perception improves 3D understanding by fusing multi-agent observations, yet intermediate-feature sharing faces strict bandwidth constraints as dense BEV features saturate V2X links. We observe that collaborators view the same physical world, making their features strongly correlated; thus receivers only need innovation beyond their local context. Revisiting this from a distributed source coding perspective, we propose V2X-DSC, a framework with a Conditional Codec (DCC) for bandwidth-constrained fusion. The sender compresses BEV features into compact codes, while the receiver performs conditional reconstruction using its local features as side information, allocating bits to complementary cues rather than redundant content. This conditional structure regularizes learning, encouraging incremental representation and yielding lower-noise features. Experiments on DAIR-V2X, OPV2V, and V2X-Real demonstrate state-of-the-art accuracy-bandwidth trade-offs under KB-level communication, and generalizes as a plug-and-play communication layer across multiple fusion backbones.

Figures (4)

• Figure 1: Overall pipeline of V2X-DSC with the proposed DSC-guided Conditional Codec (DCC). Each agent extracts a BEV feature map $\mathbf{F}_k=\Phi(\mathcal{O}_k)$ from its local observation. For a directed interaction $j\rightarrow i$, the sender prunes task-irrelevant regions to obtain $\tilde{\mathbf{F}}_j$, encodes it into a compressed bitstream $m_{j\rightarrow i}$, and transmits it to the receiver. The receiver decodes $m_{j\rightarrow i}$ and reconstructs a fusion-ready feature $\hat{\mathbf{F}}_{j\rightarrow i}$ conditioned on its local feature $\mathbf{F}_i$ (side information). Reconstructed features are fused with $\mathbf{F}_i$ to form $\mathbf{F}^{\mathrm{fused}}_i$, which is fed to the task head to produce the final prediction $\hat{\mathbf{y}}_i$. Color coding: blue indicates sender-side processing and transmitted representations, while green indicates receiver-side processing that leverages locally available information.
• Figure 2: Architecture of the DSC-guided Conditional Codec (DCC). Given the sender feature $\mathbf{F}_j$, the source-information encoder maps it to a compact latent representation, which is discretized by codebook quantization (Q). The resulting discrete symbols are entropy-coded into a compressed bitstream via rANS encoding (RE) and transmitted over the communication channel. On the receiver side, rANS decoding (RD) and dequantization (DQ) recover the quantized latent. In parallel, the receiver-local feature $\mathbf{F}_i$ is processed by a side-information encoder to produce conditioning features. The conditional decoder then combines the decoded latent with the side-information branch and reconstructs a fusion-ready feature $\hat{\mathbf{F}}_{j\rightarrow i}$.
• Figure 3: Communication--performance curves under different bandwidth budgets. The x-axis shows the average per-link payload measured in bytes and plotted in $\log_2$ scale. We vary the effective communication rate by adjusting the pruning ratio, and report detection performance under the corresponding bandwidth. V2X-DSC achieves a favorable trade-off and remains on the Pareto frontier, particularly in the low-bandwidth regime.
• Figure 4: Robustness analysis on the OPV2V dataset. (a) Impact of Localization Noise: We evaluate detection performance under increasing Gaussian noise added to relative poses. V2X-DSC exhibits relatively stable performance compared to baselines as noise intensity increases. (b) Impact of Communication Latency: Performance comparison under varying frame delays. Our method exhibits gradual performance degradation rather than a steep decline as latency increases.