Feature Pyramid Networks (FPN) are introduced to efficiently build multi-scale feature pyramids within deep convolutional networks (ConvNets) for object detection. The FPN architecture combines low-resolution, semantically strong features with high-resolution, semantically weak features through a top-down pathway and lateral connections, enabling the creation of a feature pyramid with strong semantics at all scales. This approach avoids the computational and memory costs of traditional image pyramids while maintaining high accuracy. The FPN is implemented using a top-down architecture with lateral connections, which allows for the generation of high-level semantic feature maps at all scales. This architecture is applied within the Faster R-CNN framework, achieving state-of-the-art results on the COCO detection benchmark without additional bells and whistles. The method runs at 6 FPS on a GPU, making it both practical and accurate for multi-scale object detection. The FPN is evaluated on various tasks, including object detection and segmentation. It significantly improves performance on small objects and enhances the accuracy of object detection compared to single-scale baselines. The method is also extended to instance segmentation, where it generates segmentation proposals using a similar approach to object detection. The FPN is shown to be effective in replacing traditional image pyramids for multi-scale detection tasks, offering a generic feature extractor that is both efficient and accurate. The results demonstrate that FPNs achieve higher accuracy than existing state-of-the-art methods without increasing testing time. The method is implemented in a fully convolutional manner, allowing for end-to-end training and efficient inference. The FPN has been shown to outperform previous methods in object detection and segmentation, and it is a practical solution for multi-scale detection problems.Feature Pyramid Networks (FPN) are introduced to efficiently build multi-scale feature pyramids within deep convolutional networks (ConvNets) for object detection. The FPN architecture combines low-resolution, semantically strong features with high-resolution, semantically weak features through a top-down pathway and lateral connections, enabling the creation of a feature pyramid with strong semantics at all scales. This approach avoids the computational and memory costs of traditional image pyramids while maintaining high accuracy. The FPN is implemented using a top-down architecture with lateral connections, which allows for the generation of high-level semantic feature maps at all scales. This architecture is applied within the Faster R-CNN framework, achieving state-of-the-art results on the COCO detection benchmark without additional bells and whistles. The method runs at 6 FPS on a GPU, making it both practical and accurate for multi-scale object detection. The FPN is evaluated on various tasks, including object detection and segmentation. It significantly improves performance on small objects and enhances the accuracy of object detection compared to single-scale baselines. The method is also extended to instance segmentation, where it generates segmentation proposals using a similar approach to object detection. The FPN is shown to be effective in replacing traditional image pyramids for multi-scale detection tasks, offering a generic feature extractor that is both efficient and accurate. The results demonstrate that FPNs achieve higher accuracy than existing state-of-the-art methods without increasing testing time. The method is implemented in a fully convolutional manner, allowing for end-to-end training and efficient inference. The FPN has been shown to outperform previous methods in object detection and segmentation, and it is a practical solution for multi-scale detection problems.