The paper introduces Feature Pyramid Networks (FPNs) for object detection, leveraging the inherent multi-scale hierarchy of deep convolutional networks to construct feature pyramids with minimal extra computational cost. FPNs are designed to improve the performance of object detectors by providing high-level semantic features at multiple scales. The architecture combines lateral connections to merge low-resolution, semantically strong features with high-resolution, semantically weak features, creating a rich and accurate feature pyramid. This approach is evaluated on the COCO dataset, achieving state-of-the-art results without additional engineering, surpassing all existing single-model entries, including those from the COCO 2016 challenge winners. FPNs are also shown to be efficient, running at 6 FPS on a GPU, making them a practical solution for multi-scale object detection. The method is applicable to various tasks, including instance segmentation and keypoint estimation, and the code is made publicly available.The paper introduces Feature Pyramid Networks (FPNs) for object detection, leveraging the inherent multi-scale hierarchy of deep convolutional networks to construct feature pyramids with minimal extra computational cost. FPNs are designed to improve the performance of object detectors by providing high-level semantic features at multiple scales. The architecture combines lateral connections to merge low-resolution, semantically strong features with high-resolution, semantically weak features, creating a rich and accurate feature pyramid. This approach is evaluated on the COCO dataset, achieving state-of-the-art results without additional engineering, surpassing all existing single-model entries, including those from the COCO 2016 challenge winners. FPNs are also shown to be efficient, running at 6 FPS on a GPU, making them a practical solution for multi-scale object detection. The method is applicable to various tasks, including instance segmentation and keypoint estimation, and the code is made publicly available.