Fast R-CNN

Fast R-CNN

罗斯·吉尔希克 Ross Girshick · Microsoft · 2015-04-30 · arXiv:1504.08083 ↗ · 被引 28618

打开互动全文版（逐段中英对照 + 图/公式 + 论文问答）→

摘要 · Abstract

本文提出了一种基于区域的快速卷积网络方法（Fast R-CNN）用于目标检测。Fast R-CNN 在先前工作的基础上，利用深度卷积网络高效地对目标提议进行分类。与先前工作相比，Fast R-CNN 采用多项创新，提高了训练和测试速度，同时提升了检测精度。Fast R-CNN 训练非常深的 VGG16 网络比 R-CNN 快 9 倍，测试时快 213 倍，并在 PASCAL VOC 2012 上获得了更高的 mAP。与 SPPnet 相比，Fast R-CNN 训练 VGG16 快 3 倍，测试快 10 倍，且更准确。Fast R-CNN 使用 Python 和 C++（基于 Caffe）实现，并在 MIT 开源许可下于 https://github.com/rbgirshick/fast-rcnn 提供。

This paper proposes a Fast Region-based Convolutional Network method (Fast R-CNN) for object detection. Fast R-CNN builds on previous work to efficiently classify object proposals using deep convolutional networks. Compared to previous work, Fast R-CNN employs several innovations to improve training and testing speed while also increasing detection accuracy. Fast R-CNN trains the very deep VGG16 network 9x faster than R-CNN, is 213x faster at test-time, and achieves a higher mAP on PASCAL VOC 2012. Compared to SPPnet, Fast R-CNN trains VGG16 3x faster, tests 10x faster, and is more accurate. Fast R-CNN is implemented in Python and C++ (using Caffe) and is available under the open-source MIT License at https://github.com/rbgirshick/fast-rcnn.

核心贡献 · Key contributions

• 提出 Fast R-CNN，一种用于目标检测的单阶段训练算法，联合学习分类和边界框回归。
  Proposes Fast R-CNN, a single-stage training algorithm for object detection that jointly learns classification and bounding-box regression.
• 在 VGG16 上实现比 R-CNN 快 9 倍的训练速度和 213 倍的测试速度，同时在 PASCAL VOC 2012 上提高了 mAP。
  Achieves 9x faster training and 213x faster testing than R-CNN with VGG16, while improving mAP on PASCAL VOC 2012.
• 引入 RoI 池化层，通过区域提议实现高效的特征共享和反向传播。
  Introduces RoI pooling layer for efficient feature sharing and back-propagation through region proposals.
• 使用多任务损失，结合 softmax 分类和鲁棒的 L1 边界框回归。
  Uses multi-task loss combining softmax classification and robust L1 bounding-box regression.
• 证明微调卷积层对于 VGG16 等非常深的网络至关重要。
  Demonstrates that fine-tuning convolutional layers is crucial for very deep networks like VGG16.
• 表明稀疏目标提议可提高检测器质量，实现高效评估。
  Shows that sparse object proposals improve detector quality, enabling efficient evaluation.

局限 · Limitations

• 依赖预计算的目标提议，限制了端到端训练，可能遗漏目标。
  Relies on pre-computed object proposals, which limits end-to-end training and may miss objects.
• 单尺度训练和测试可能无法很好地泛化到极端尺度变化。
  Single-scale training and testing may not generalize well to extreme scale variations.
• GPU 内存限制阻碍了多尺度训练中使用非常深的网络。
  GPU memory constraints limit the use of very deep networks with multi-scale training.
• 使用密集目标提议而非稀疏提议时性能下降。
  Performance degrades when using dense object proposals instead of sparse ones.
• 方法仅在 PASCAL VOC 和 MS COCO 数据集上评估，未测试对其他领域的泛化性。
  Method is evaluated only on PASCAL VOC and MS COCO datasets; generalization to other domains is not tested.

论文章节 · Sections（共 25）

• 摘要 Abstract
• 1 引言 1 Introduction
• 1.1 R-CNN 与 SPPnet 1.1 R-CNN and SPPnet
• 1.2 贡献 1.2 Contributions
• 2 Fast R-CNN 架构与训练 2 Fast R-CNN architecture and training
• 2.1 RoI 池化层 2.1 The RoI pooling layer
• 2.2 从预训练网络初始化 2.2 Initializing from pre-trained networks
• 2.3 检测微调 2.3 Fine-tuning for detection
• 2.4 尺度不变性 2.4 Scale invariance
• 3 Fast R-CNN 检测 3 Fast R-CNN detection
• 3.1 截断 SVD 加速检测 3.1 Truncated SVD for faster detection
• 4 主要结果 4 Main results
• 4.1 实验设置 4.1 Experimental setup
• 4.2 VOC 2010 与 2012 结果 4.2 VOC 2010 and 2012 results
• 4.3 VOC 2007 结果 4.3 VOC 2007 results
• 4.4 训练与测试时间 4.4 Training and testing time
• 4.5 微调哪些层？ 4.5 Which layers to fine-tune?
• 5 设计评估 5 Design evaluation
• 5.1 多任务训练是否有帮助？ 5.1 Does multi-task training help?
• 5.2 尺度不变性：暴力还是技巧？ 5.2 Scale invariance: to brute force or finesse?
• 5.3 是否需要更多训练数据？ 5.3 Do we need more training data?
• 5.4 SVM 是否优于 softmax？ 5.4 Do SVMs outperform softmax?
• 5.5 更多候选框是否总是更好？ 5.5 Are more proposals always better?
• 5.6 初步 MS COCO 结果 5.6 Preliminary MS COCO results
• 6 结论 6 Conclusion

阅读逐段中英对照全文 →