最先进的目标检测网络依赖于区域提议算法来假设目标位置。SPPnet 和 Fast R-CNN 等进展减少了这些检测网络的运行时间，但区域提议计算成为瓶颈。在这项工作中，我们引入了一个区域提议网络（RPN），它与检测网络共享全图像卷积特征，从而实现了几乎无成本的区域提议。RPN 是一个全卷积网络，同时预测每个位置的目标边界和目标性得分。RPN 经过端到端训练以生成高质量的区域提议，供 Fast R-CNN 用于检测。我们进一步将 RPN 和 Fast R-CNN 合并为一个单一网络，通过共享它们的卷积特征——使用最近流行的神经网络“注意力”机制术语，RPN 组件告诉统一网络在哪里看。对于非常深的 VGG-16 模型，我们的检测系统在 GPU 上达到 5fps 的帧率（包括所有步骤），同时在 PASCAL VOC 2007、2012 和 MS COCO 数据集上仅使用每张图像 300 个提议就实现了最先进的目标检测精度。在 ILSVRC 和 COCO 2015 竞赛中，Faster R-CNN 和 RPN 是多个赛道第一名获奖作品的基础。代码已公开。
State-of-the-art object detection networks depend on region proposal algorithms to hypothesize object locations. Advances like SPPnet and Fast R-CNN have reduced the running time of these detection networks, exposing region proposal computation as a bottleneck. In this work, we introduce a Region Proposal Network (RPN) that shares full-image convolutional features with the detection network, thus enabling nearly cost-free region proposals. An RPN is a fully convolutional network that simultaneously predicts object bounds and objectness scores at each position. The RPN is trained end-to-end to generate high-quality region proposals, which are used by Fast R-CNN for detection. We further merge RPN and Fast R-CNN into a single network by sharing their convolutional features---using the recently popular terminology of neural networks with 'attention' mechanisms, the RPN component tells the unified network where to look. For the very deep VGG-16 model, our detection system has a frame rate of 5fps (including all steps) on a GPU, while achieving state-of-the-art object detection accuracy on PASCAL VOC 2007, 2012, and MS COCO datasets with only 300 proposals per image. In ILSVRC and COCO 2015 competitions, Faster R-CNN and RPN are the foundations of the 1st-place winning entries in several tracks. Code has been made publicly available.
核心贡献 · Key contributions
提出区域提议网络（RPN），与检测网络共享卷积特征，实现近乎零成本的区域提议。 Introduces Region Proposal Network (RPN) sharing convolutional features with detection network, enabling nearly cost-free region proposals.
提出锚点框作为回归参考，无需图像或滤波器金字塔即可实现多尺度和长宽比预测。 Proposes anchor boxes as regression references for multi-scale and aspect ratio prediction without image or filter pyramids.
开发四步交替训练算法，使 RPN 和 Fast R-CNN 共享卷积层，形成统一网络。 Develops a 4-step alternating training algorithm to share convolutional layers between RPN and Fast R-CNN, forming a unified network.
在 PASCAL VOC 2007、2012 和 MS COCO 上，仅用每图 300 个提议即达到最先进检测精度。 Achieves state-of-the-art detection accuracy on PASCAL VOC 2007, 2012, and MS COCO with only 300 proposals per image.
在 VGG-16 上实现近实时性能，GPU 上 5fps，包含所有检测步骤。 Demonstrates near real-time performance at 5fps on GPU with VGG-16, including all detection steps.
作为 ILSVRC 和 COCO 2015 竞赛多个赛道第一名的基础。 Serves as foundation for 1st-place entries in ILSVRC and COCO 2015 competitions across multiple tracks.
局限 · Limitations
RPN 依赖预定义的锚点尺度和长宽比，可能无法泛化到所有物体形状。 RPN relies on pre-defined anchor scales and aspect ratios, which may not generalize to all object shapes.
训练需谨慎处理跨边界锚点以避免收敛问题。 Training requires careful handling of cross-boundary anchors to avoid convergence issues.
需非极大值抑制减少提议冗余，增加后处理步骤。 Non-maximum suppression is needed to reduce proposal redundancy, adding a post-processing step.
性能依赖 ImageNet 预训练模型，限制了在数据分布不同领域的适用性。 Performance depends on pre-trained ImageNet models, limiting applicability to domains with different data distributions.
RPN 和 Fast R-CNN 的联合训练是近似的，忽略了关于提议框坐标的梯度。 Joint training of RPN and Fast R-CNN is approximate, ignoring gradients w.r.t. proposal box coordinates.
论文章节 · Sections（共 11）
摘要Abstract
1 引言1 Introduction
2 相关工作2 Related Work
3 更快的 R-CNN3 Faster R-CNN
3.1 区域提议网络3.1 Region Proposal Networks
3.2 共享 RPN 和 Fast R-CNN 的特征3.2 Sharing Features for RPN and Fast R-CNN
3.3 实现细节3.3 Implementation Details
4.1 PASCAL VOC 实验4.1 Experiments on PASCAL VOC
4.2 MS COCO 实验4.2 Experiments on MS COCO
4.3 从 MS COCO 到 PASCAL VOC4.3 From MS COCO to PASCAL VOC