Stair-Step Feature Pyramid Networks for Object Detection

Abstract

Feature Pyramid Networks have solved scale variation problems in object detection by developing multi-level features with different scales from backbone networks. Although this network achieved promising performance without affecting model complexity, they still suffer feature-level imbalance between multi-level features, i.e., low-level features and high-level features in each stage of the backbone. Moreover, the detection head predicts classification scores and offset regression independently on each feature of multi-level features, which leads to inconsistency among the detection branch. Hence, this paper releases this problem by introducing simple but effective Stair-step Feature Pyramid Networks (SFPN) to harmonize information between multi-level features. Further, the Offset Adaption Module (OA Module) is proposed to improve feature representation by adapting the feature of the classification branch with regressed offsets of the regression branch. On the MS-COCO dataset, the proposed method increases by 1.2% Average Precision when comparing with baseline FCOS without bells and whistles.