A weighted converging and radiating feature pyramid network for object detection

Contemporary state-of-the-art detectors typically employ multi-scale features to detect objects at varying scales. A common approach for extracting multi-scale features is utilizing a feature pyramid network, implemented by fusing features through one-way summation or concatenation. As a result, they cannot take into account the information from both the upper and lower adjacent layers simultaneously, thus impairing the representation of multi-scale features. In this paper, we first propose a converging and radiating pathway to enhance the representation of features by fusing features from the adjacent upper and lower layers directly. Second, to optimize the feature fusion in the proposed pyramid pathway, we replace the conventional summation operation in the fusion process with weighted summation, adaptive channel fusion (ACF), and adaptive spatial fusion (ASF) operations. Finally, a weighted converging and radiating feature pyramid network (CRFPN) is developed based on the above-described methods, achieving more competitive results than other state-of-the-art feature pyramid networks. Furthermore, we have integrated the proposed CRFPN into various detectors and backbones and experimentally verified the adaptability and generalizability of our method. The code is available at https://github.com/gmpan66/CRFPN.