用于精确目标检测和语义分割的丰富特征层次结构

Rich feature hierarchies for accurate object detection and semantic segmentation

罗斯·吉尔希克 Ross Girshick · UC Berkeley · 2013-11-11 · arXiv:1311.2524 ↗ · 被引 29100

摘要 · Abstract

在标准 PASCAL VOC 数据集上，目标检测性能在过去几年中停滞不前。表现最好的方法是复杂的集成系统，通常将多个低级图像特征与高级上下文相结合。在本文中，我们提出了一种简单且可扩展的检测算法，相比 VOC 2012 上之前的最佳结果，平均精度均值（mAP）相对提高了 30%以上，达到了 53.3%的 mAP。我们的方法结合了两个关键见解：（1）可以将高容量卷积神经网络（CNN）应用于自底向上的区域提议，以定位和分割目标；（2）当标记训练数据稀缺时，对辅助任务进行监督预训练，然后进行领域特定的微调，可以显著提升性能。由于我们将区域提议与 CNN 结合，我们将我们的方法称为 R-CNN：具有 CNN 特征的区域。我们还将 R-CNN 与 OverFeat（一种基于类似 CNN 架构的最近提出的滑动窗口检测器）进行了比较。我们发现，在 200 类的 ILSVRC2013 检测数据集上，R-CNN 大幅优于 OverFeat。完整系统的源代码可在 http://www.cs.berkeley.edu/~rbg/rcnn 获取。

Object detection performance, as measured on the canonical PASCAL VOC dataset, has plateaued in the last few years. The best-performing methods are complex ensemble systems that typically combine multiple low-level image features with high-level context. In this paper, we propose a simple and scalable detection algorithm that improves mean average precision (mAP) by more than 30% relative to the previous best result on VOC 2012---achieving a mAP of 53.3%. Our approach combines two key insights: (1) one can apply high-capacity convolutional neural networks (CNNs) to bottom-up region proposals in order to localize and segment objects and (2) when labeled training data is scarce, supervised pre-training for an auxiliary task, followed by domain-specific fine-tuning, yields a significant performance boost. Since we combine region proposals with CNNs, we call our method R-CNN: Regions with CNN features. We also compare R-CNN to OverFeat, a recently proposed sliding-window detector based on a similar CNN architecture. We find that R-CNN outperforms OverFeat by a large margin on the 200-class ILSVRC2013 detection dataset. Source code for the complete system is available at http://www.cs.berkeley.edu/~rbg/rcnn.

核心贡献 · Key contributions

提出 R-CNN，将区域提议与 CNN 结合用于目标检测。
Proposes R-CNN, combining region proposals with CNNs for object detection.
在 PASCAL VOC 2012 上相比先前最佳结果实现 30%的相对 mAP 提升。
Achieves 30% relative mAP improvement on PASCAL VOC 2012 over prior best.
展示在辅助任务上有监督预训练加领域特定微调可提升性能。
Shows supervised pre-training on auxiliary task plus domain-specific fine-tuning boosts performance.
证明 R-CNN 在 ILSVRC2013 检测上大幅优于 OverFeat。
Demonstrates R-CNN outperforms OverFeat on ILSVRC2013 detection by large margin.
将 R-CNN 扩展到语义分割，在 PASCAL VOC 上取得有竞争力的结果。
Extends R-CNN to semantic segmentation with competitive results on PASCAL VOC.
提供错误模式分析和边界框回归以减少定位错误。
Provides analysis of error modes and bounding-box regression to reduce mislocalizations.

局限 · Limitations

需要通过扭曲获得固定大小的 CNN 输入，可能扭曲宽高比。
Requires fixed-size CNN input via warping, which may distort aspect ratios.
区域提议阶段（选择性搜索）是速度和召回率的瓶颈。
Region proposal stage (selective search) is a bottleneck for speed and recall.
微调和 SVM 训练使用不同的正负样本定义，导致不一致。
Fine-tuning and SVM training use different positive/negative definitions, causing inconsistency.
边界框回归仅对接近真实值的提议有效，限制了适用范围。
Bounding-box regression only effective for proposals near ground truth, limiting scope.
性能严重依赖预训练数据集（ImageNet）和架构选择。
Performance depends heavily on pre-training dataset (ImageNet) and architecture choice.

论文章节 · Sections（共 29）

摘要 Abstract
1 引言 1 Introduction
2 基于 R-CNN 的目标检测 2 Object detection with R-CNN
2.1 模块设计 2.1 Module design
2.2 测试时检测 2.2 Test-time detection
2.3 训练 2.3 Training
2.4 在 PASCAL VOC 2010-12 上的结果 2.4 Results on PASCAL VOC 2010-12
2.5 在 ILSVRC2013 检测上的结果 2.5 Results on ILSVRC2013 detection
3.1 可视化学习特征 3.1 Visualizing learned features
3.2 消融研究 3.2 Ablation studies
3.3 网络架构 3.3 Network architectures
3.4 检测错误分析 3.4 Detection error analysis
3.5 边界框回归 3.5 Bounding-box regression
3.6 定性结果 3.6 Qualitative results
4 ILSVRC2013 检测数据集 4 The ILSVRC2013 detection dataset
4.1 数据集概述 4.1 Dataset overview
4.2 区域提议 4.2 Region proposals
4.3 训练数据 4.3 Training data
4.4 验证与评估 4.4 Validation and evaluation
4.6 与 OverFeat 的关系 4.6 Relationship to OverFeat
5 语义分割 5 Semantic segmentation
6 结论 6 Conclusion
附录 A 目标提议变换 Appendix A Object proposal transformations
附录 B 正负样本与 Softmax Appendix B Positive vs. negative examples and softmax
附录 C 边界框回归 Appendix C Bounding-box regression
附录 D 额外特征可视化 Appendix D Additional feature visualizations
附录 E 每类分割结果 Appendix E Per-category segmentation results
附录 F 跨数据集冗余分析 Appendix F Analysis of cross-dataset redundancy
附录 G 文档变更日志 Appendix G Document changelog

阅读逐段中英对照全文 →