DCFNet: A Distinction-aware Cross-stage Fusion Network for RGB-T Crowd Counting

RGB-T crowd counting gains robustness against poor illumination conditions by virtue of the complementarity of multimodal data. Although most existing methods have demonstrated promising results in fusing multimodal data, they suffer from two drawbacks. First, they overlook the information distinctions between different modalities, which is not conducive to exploring complementary information, thus potentially impeding fusion feature learning. Second, crowd information hidden in the hierarchical features is not fully mined, which makes it challenging to generate excellent density maps with limited feature representations. To overcome these drawbacks, we propose a Distinction-aware Cross-stage Fusion Network, namely DCFNet. It could effectively leverage information distinctions between modalities to preserve more complementary information and sufficiently delve into the hidden hierarchical information. Specifically, our DCFNet mainly comprises the Distinction-aware Multimodal Fusion (DMF) module and the Cross-stage Information Enhancement (CIE) module. The DMF module focuses on learning distinct information across modalities to raise the discriminable ability of fusion features and also provides a novel perspective for RGB-T crowd counting about how to explore valuable complementary information. Simultaneously, the CIE module aims to aggregate the cross-stage features and extract the maximum multi-level crowd information hidden in the hierarchical features, thereby enhancing the representation ability of the learned features. The DCFNet has been evaluated on two mainstream benchmarks, demonstrating its effectiveness and SOTA counting performance.