HAFNet: Hierarchical Attentive Fusion Network for Multispectral Pedestrian Detection

Multispectral pedestrian detection via visible and thermal image pairs has received widespread attention in recent years. It provides a promising multi-modality solution to address the challenges of low-light environments occlusion situations. Most existing methods directly blend results two modalities or combine features linear interpolation. However, such fusion strategies tend extract coarser corresponding positions different modalities, which may lead degraded performance. To mitigate this, this paper proposes novel adaptive cross-modality framework, named Hierarchical Attentive Fusion Network (HAFNet), fully exploits multispectral knowledge inspire decision-making process. Concretely, we introduce Content-dependent (HCAF) module top-level as guide pixel-wise blending enhance quality feature representation plug-in alignment (MFA) block fine-tune modalities. Experiments on challenging KAIST CVC-14 datasets demonstrate superior performance our method with satisfactory speed.