Accurate 3-D object detection in cooperative vehicle-to-infrastructure (V2X) scenarios presents significant challenges due to heterogeneous sensor modalities, semantic inconsistency across domains, and the need for efficient information transmission in real time. To address these issues, we propose CoFeatNet, a unified and lightweight multimodal feature extraction network tailored for cooperative perception tasks. CoFeatNet integrates depthwise separable convolutions, adaptive pyramid sampling, transposed convolutions, and both channel-wise and cross-scale attention mechanisms into a cohesive framework that effectively reduces structural discrepancies among heterogeneous features, enhances semantic interactions between vehicle- and infrastructure-side inputs, and minimizes computational overhead. In addition, a geometry-aware compensation mechanism is introduced to mitigate distortion effects in monocular imagery by leveraging intrinsic camera parameters through an attention-guided strategy. Extensive experiments conducted on the DAIR-V2X dataset demonstrate the effectiveness of our approach, with CoFeatNet achieving 17.40% AP${}_{\text {3-D}}$ and 23.92% APBEV, outperforming the EMIFF baseline by 1.79% and 2.48%, respectively. These results highlight the superior accuracy and efficiency of CoFeatNet in complex V2X cooperative environments.