Information-Interaction Feature Pyramid Networks for Object Detection

Abstract

Information interaction between multi-scale features is crucial for recognition systems detecting objects at different scales. In this paper, an Information-Interaction Feature Pyramid Network (IFPN) is proposed to enhance the power of the entire feature representations in a simple but efficient way. Specifically, to strengthen the longitudinal information interaction between multi-scale features, we establish a Bidirectional Information Pyramid Network, which significantly enhances all level features with reasonable localization and classification capabilities. Furthermore, Residual Information Branches are constructed to optimize the lateral information flow between the input and output neurons of the same middle pyramid levels. Taking Feature Pyramid Network (FPN) as the benchmark, by replacing Path Aggregation Network (PANet) with IFPN, our method achieves 3.5x and 1.6x Average Precision (AP) improvement in Faster R-CNN and YOLOX-Nano, respectively. With higher accuracy, IFPN uses 15% fewer GFLOPs than the Balanced Feature Pyramid (BFP) in YOLOX-Nano, achieving better speed and accuracy trade-offs. Furthermore, when IFPN replaces FPN, our method improves Mask R-CNN by 1.1% AP and RetinaNet by 1.0% AP, respectively, when using ResNet-50 as the backbone.