One camera at each end of a V2I link: cross-image routing attention lifts top-1 beam-pair accuracy by 3.15 points

Millimeter-wave V2I beam-pair power prediction no longer relies on a camera at only one end of the link. CiFormer uses paired vehicle-side and roadside RGB images to output a power matrix over all candidate beam pairs and to select the best beam pair from it. Earlier sensing-aided beam prediction methods mainly used observations from either the vehicle or the infrastructure, capturing only a partial view of the propagation environment and leaving link-relevant information across the two endpoints underexploited. On the DuSen-BP3 dataset, CiFormer reaches 72.41% top-1 beam-pair selection accuracy, 3.15 percentage points above the strongest baseline MambaOut at 69.26%, and 94.68% top-3 accuracy, 1.65 points above MambaOut.

The two images, taken from very different viewpoints, are first split into windows, and each window is linked only to the few most relevant windows in the other view before token-level attention is computed inside those windows. Cross-image routing attention (CIRA) extends BiFormer's bi-level routing from within a single feature map to between two images, using a learnable top-k route to pick the most relevant windows in the opposite view instead of attending densely to all regions. Ablations show RMSE rising from 5.810 dB to 6.001 dB without cross-view interaction, to 5.922 dB without top-k routing, and to 5.893 dB without the semantic-aware bias, indicating that selective cross-view association contributes to prediction.

The report covers mmWave V2I beam-pair power prediction: the input is one RGB image from each end of the link, and the output is a power matrix over all candidate beam pairs from which the best pair is selected. Experiments run in a 28 GHz V2I simulation with a uniform planar array and a seven-beam codebook at each endpoint; the DuSen-BP3 dataset contains 6,965 valid samples split 7:1:2 into training, validation and test sets. The authors suggest the mechanism could extend to channel modeling, beam management and resource allocation, but those directions are not validated here.

Results come from the authors' simulated DuSen-BP3 dataset with six vehicles, five base stations and seven links, and have not been validated on real-road measurements. Ablations show RMSE rises more when the roadside view is removed, indicating greater sensitivity to receiver-side information; the Grad-CAM visualizations are offered as qualitative support rather than causal evidence. The authors also note that applicability of cross-view interaction to broader wireless tasks remains speculative.