DinoLink: A Token-Centric Representation Compression Framework for Bandwidth-Constrained Collaborative V2X Perception

Autonomous driving systems increasingly rely on Vehicle-to-Everything (V2X) communication to offload heavy perception tasks to cloud servers. However, current methods are bottlenecked by bandwidth constraints. Standard image codecs like JPEG are optimized for human vision, often discarding the fine-grained semantic details required by machine vision models. Conversely, transmitting raw neural feature maps is too data-intensive for real-world wireless networks. DinoLink addresses this by shifting from pixel-level streaming to a discrete, semantic-centric communication paradigm.

DinoLink employs a dual-sparsity architecture to minimize bandwidth usage while preserving task-critical information. First, a Saliency-Aware Token Selector uses DINOv2 attention maps to identify and retain only the most informative spatial tokens, discarding redundant background regions like sky or empty road. Second, the selected tokens are processed through a Residual Vector Quantization (RVQ) module, which maps continuous high-dimensional features into a sequence of compact, discrete codebook indices. The system transmits only these lightweight indices and their positional priors, effectively decoupling the compression interface from the specific downstream perception model.

Evaluations on the nuScenes dataset demonstrate that DinoLink achieves a 139x reduction in bitrate compared to uncompressed transmission, while maintaining a competitive 32.8% mAP. In simulated narrow-band environments, such as LoRa, the framework provides a 34.5x acceleration in data transmission. By replacing dense feature maps with sparse, quantized tokens, DinoLink enables high-fidelity remote perception even under the volatile and strictly constrained conditions typical of V2X networks.