Communication Efficient Cooperative Perception via Codebook-Free Vector Quantization

Recent advances in cooperative perception have demonstrated significant performance improvements over single-agent perception. In practice, cooperative perception methods often exchange intermediate multi-channel BEV feature maps at a high frame rate, which can impose a substantial burden on available V2X bandwidth and hinder the practical deployment of such systems. To mitigate bandwidth limitations, we introduce Codebook-Free Quantization for Cooperative Perception (CFQ4CP), a lightweight feature-compression framework. Specifically, a learned neural network, combined with a simple rounding operation, maps continuous latent features to compact integer codes. The discrete representations, instead of high-dimensional feature maps, are then transmitted through V2X communication. Subsequently, these features are reconstructed from the integer codes via a lightweight neural network compression decoder. Unlike existing compression methods that rely on fixed codebooks and nearest-neighbor lookups, our approach generates highly compact integer representations that eliminate the need for codebook alignment between transmitting and receiving agents. We evaluated the proposed framework on two real-world cooperative perception datasets, V2V4Real and DAIR-V2X. Experimental results demonstrate that CFQ4CP achieves perception accuracy competitive with full-precision models, even when each feature vector in the bird’s eye view map is quantized to extremely low bitwidths (e.g., 1 or 2 bits), results in reducing the communication costs by thousands of times.