Common Extraction and Distribution Guided for Weakly Supervised RGB-IR Vehicle Detection

Nowadays, object detection on RGB-Infrared (IR) natural images has achieved remarkable success. However, compared to natural RGB-IR images, aerial RGB-IR images suffer from more serious local feature failure because of imbalanced illumination. Though in this area, most existing methods ignore local feature failure because of imbalanced illumination distribution in aerial images. Therefore, we attempt to use illumination distribution to adaptively guide the feature fusion process, thereby reducing the influence of local feature failure. Additionally, leveraging the spatial correlation between cross-modal feature maps allows for the dynamic selection of common feature information. In this work, a Common Extraction and Distribution Guided Network (CEDGNet) is presented for aerial RGB-IR images that can accurately detect oriented objects even when local features failure and horizontal box (HBox) annotations. Specifically, a Distribution Effective Guidance (DEG) module is designed to guide the fusion process and eliminate the interference of local failure features because of imbalance illumination. Furthermore, a Common Feature Extraction (CFE) module is proposed to enhance the effective features between modalities by obtaining common attention information correlated between modalities. Comprehensive experiments on the public DroneVehicle datasets demonstrate that our method reduces the effect of the local feature failure and CEDGNet outperforms relevant state-of-the-art algorithms.