CAC-Net: Convolutional Additive Cross-Modal Fusion for Real-Time RGB-T Object Detection
Ziwen Zhu, Xiaoou Song, Guangming Zhou, Mingmei Chen, Xiaobing Deng · IEEE Access · 2026
Real-time RGB-T object detection faces significant challenges from illumination variations, background thermal clutter, and strict computational constraints. To tackle these, we propose CAC-Net, an efficient convolutional additive cross-modal fusion network that leverages complementary visible and thermal infrared modalities while maintaining linear computational complexity in the fusion pipeline. Our key contributions are threefold: (1) integrating Pinwheel-shaped convolution in the early dual-stream backbone to enhance discriminative representation of faint thermal cues and fine-grained RGB textures; (2) enabling global dependency modeling via a cross-modal additive attention fusion module with explicit spatial and channel alignment; (3) introducing a consensus-driven shared straight-through Top-K sparsification mechanism that adaptively suppresses modality-specific noise while remaining differentiable. Extensive experiments on challenging RGB-T benchmarks demonstrate that CAC-Net significantly outperforms current publicly available state-of-the-art methods in both accuracy and efficiency, establishing a new paradigm for robust, real-time, and resource-efficient multimodal perception. The code will be available at https://github.com/ZZWHLL/CAC-Net.