Weighted parallel decoupled feature pyramid network for object detection

Abstract

As one of the most representative detectors, feature pyramid network (FPN) has achieved remarkable improvement in object detection. Thanks to its distinctive top-down feature fusion path and multi-scale detection paradigm, FPN has become an essential component in modern detectors and has attracted increasing attention. Nevertheless, since the information of each level feature is not object-specific and the heterogeneity of classification and localization tasks in detection, FPN and its numerous variants endure information redundancy and task conflict. To address these two deficiencies that greatly limit the detection performance, this paper proposes weighted parallel decoupled FPN (WPDFPN). Specifically, selective fusion and elimination (SFE) focusing on internal construction of pyramid is firstly proposed. Compared to existing feature fusion methods, SFE effectively fuses complementary information between different level features and eliminates redundant information to produce object-specific feature at each pyramid level. Then, a novel weighted parallel FPN (WPFPN) is constructed by investigating the combination manners of feature fusion paths. Different from the sequential bidirectional pyramids, WPFPN performs the top-down and bottom-up paths in parallel and aggregates them using a weighted strategy, adaptively integrating object semantics and details. On this basis, feature decoupling mechanism and decoupled head are both developed to learn task-specific features for classification and localization, thus effectively mitigating the task conflict. Without bells and whistles, WPDFPN achieves 43.6 AP on the MS COCO dataset with ResNet-101 as backbone in Faster R-CNN framework. This achievement suppresses the baseline FPN by 3.1 points and demonstrates competitive performance in comparison with some state-of-the-art detectors. Code is available at https://github.com/HB-X/WPDFPN.