(论文阅读17/100)R-FCN: Object Detection via Region-based Fully Convolutional Networks

文献阅读笔记

简介

题目

R-FCN: Object Detection via Region-based Fully Convolutional Networks

作者

Jifeng Dai  Yi Li  Tsinghua University Kaiming He  Jian Sun

原文链接

https://arxiv.org/abs/1605.06409

详解R-FCN - 知乎 (zhihu.com)

关键词

Region-based Fully Convolutional Networks基于区域的全卷积网络

研究问题

  • For accurate and efficient Object Detection.为了更加精确和有效率的目标检测
  • Dilemma:between translation-invariance in image classfication and translation-variance in object detection. 困境:图像分类任务所要求的平移不变性和目标检测要求的平移可变性的冲突
  • 图像分类网络使用全卷积层,所以自然的,目标检测也使用全卷积层。由于以上困境导致网络结构的非自然设计:使用全卷积层的网络出现了低精度问题,所以为了解决这个问题,在两个全卷积层中非自然的插入了ROI池化层。ROI没有共享计算,所以低速。

研究方法

  • Fully convolutional with almost all computation shared on the entire image.几乎所有的计算都运行在整张图像上的全卷积层
  • It proposes Position-sensitive score maps 提出了位置敏感得分地图
  • extract candidate regions by the Region Proposal Network (RPN) [18], which is a fully convolutional architecture in itself.通过rpn网络(全卷积网络)提取区域建议。
  • On top of this FCN, we append a position-sensitive RoI pooling layer that shepherds information from these score maps, with no weight (convolutional/fc) layers following.

(论文阅读17/100)R-FCN: Object Detection via Region-based Fully Convolutional Networks_第1张图片K*k(C+1) k*k个区域 C个目标类 1个背景类

4k*k个位置信息

  • share the features between RPN and R-FCN.

(论文阅读17/100)R-FCN: Object Detection via Region-based Fully Convolutional Networks_第2张图片最后输出了每个region的得分信息

交叉熵损失函数 非极大值抑制

研究结论

  • Competitive results on the PASCAL VOC datasets with the 101-layer ResNet.83.6% mAP on the PASCAL VOC 2007 set and 82.0% the 2012 set
  • Their result is achieved at a test-time speed of 170ms per image.

额外知识

A bank of:一组;

Features map

你可能感兴趣的:(目标跟踪,人工智能,论文阅读)