ARFP: A novel adaptive recursive feature pyramid for object detection in aerial images

摘要

The aerial image is one of the most important application fields of object detection. However, the characteristics of scale variation in aerial images bring challenges to object detection. Feature pyramid network(FPN) has been proved to be a promising method for solving scale variation. However, the conventional FPN merely transfers the solid semantic information to the bottom layers through the upsampling method. Moreover, there exists no gated mechanism to distinguish effective signals. In this paper, we proposed the adaptive recursive feature pyramid(ARFP), which consists of three subparts: recursive structure, efficient global context(EGC) bottleneck module, and discriminative feature fusion(DFF) module. The DFF module makes all pyramid levels can completely leverage the firm semantic information and detailed location information through the dense connection. Besides, it adaptively learns the weights for each pyramid level to fuse and gate the signals discriminatively. The EGC module is responsible for building the pixel-wise position relevance and channel relevance, which is influential for aerial image detection. Moreover, we empirically explore the possibility of building a feature pyramid recursively. Extensive experiments on the DIOR and VisDrone2019 datasets have shown that ARFP outperforms the current state-of-art feature pyramid networks, and the average performance gain is 2.7% and 2.8%, respectively.