Hierarchical Focused Feature Pyramid Network for Small Object Detection

Small object detection has been a persistently practical and challenging task in the field of computer vision. Advanced detectors often utilize a feature pyramid network (FPN) to fuse the features generated from various receptive fields, which improve the detection ability of multi-scale objects, especially for small objects. However, existing FPNs typically employ a naive addition-based fusion strategy, which neglects crucial details that may exist only at specific levels. These details are vital for accurately detecting small objects. In this paper, we propose a novel Hierarchical Focused Feature Pyramid Network (HFFPN) to enhance these details while ensuring the detection performance for objects of other scales. HFFPN consists of two key components: Hierarchical Feature Subtraction Module (HFSM) and Feature Fusion Guidance Attention (FFGA). HFSM is first designed to selectively amplify the information important to small object detection. FFGA is devised to focus on effective features by utilizing global information and mining small objects' information from high-level features. Combining these two modules contributes greatly to the original FPN. In particular, the proposed HFFPN can be incorporated into most mainstream detectors, such as Faster RCNN, Retinanet, FCOS, etc. Extensive experiments on small object datasets demonstrate that HFFPN achieves consistent and significant improvements over the baseline algorithm while surpassing the state-of-the-art methods.