Feature Pyramid Networks for Object Detection比较FPN、UNet、Conv-Deconv
https://vitalab.github.io/deep-learning/2017/04/04/feature-pyramid-network.html
Feature Pyramid Networks for Object Detection
Reviewed on Apr 4, 2017 by Frédéric Branchaud-Charron • https://arxiv.org/pdf/1612.03144.pdf
Reference : T. Lin, P. Dollár, R. Girshick, K. He, B. Hariharan, S. Belongie. Feature Pyramid Networks for Object Detection.Computer Vision and Pattern Recognition, 2017. CVPR 2017.
The Feature Pyramid Network (FPN) looks a lot like the U-net. The main difference is that there is multiple prediction layers: one for each upsampling layer. Like the U-Net, the FPN has laterals connection between the bottom-up pyramid (left) and the top-down pyramid (right). But, where U-net only copy the features and append them, FPN apply a 1x1 convolution layer before adding them. This allows the bottom-up pyramid called “backbone” to be pretty much whatever you want. In their experiments, the authors use Resnet-50 as their backbone.

FPN for region-proposal
To achieve region-proposal, the authors add a 3x3 Conv layer followed by two 1x1 Conv for classification and regression on each upsampling layer. These additions are called heads and the weights are shared. For each head, you assign a set of anchors boxes resized to match the head’s shape. The anchors are of multiple pre-defined scales and aspect ratios in order to cover objects of different shapes. Training labels are then assigned to each anchor based on the IoU. A positive label is assigned if the IoU is greater than 70%.
FPN for object Detection
Using Fast(er) R-CNN, they can use FPN as the region proposal part. The proposals are used in combination with RoiPooling and then they can do the same work as Fast(er) R-CNN.
Results
Faster R-CNN on FPN with a ResNet-101 backbone is achieving state of the art on the COCO detection benchmark. It’s also faster than Resnet-101 Faster R-CNN by a significant margin because of the weight sharing in the heads.
Effect of lateral connections
FPN performs better than a normal Conv-Deconv because the Conv-Deconv’s feature maps are wrong according to the authors. Indeed, they argue that the locations of these maps are not precise, because these maps have been downsampled and upsampled several times. There is a 10% jump in accuracy using lateral connections.
Feature Pyramid Networks for Object Detection比较FPN、UNet、Conv-Deconv的更多相关文章
- 『计算机视觉』FPN:feature pyramid networks for object detection
对用卷积神经网络进行目标检测方法的一种改进,通过提取多尺度的特征信息进行融合,进而提高目标检测的精度,特别是在小物体检测上的精度.FPN是ResNet或DenseNet等通用特征提取网络的附加组件,可 ...
- 【Network Architecture】Feature Pyramid Networks for Object Detection(FPN)论文解析(转)
目录 0. 前言 1. 博客一 2.. 博客二 0. 前言 这篇论文提出了一种新的特征融合方式来解决多尺度问题, 感觉挺有创新性的, 如果需要与其他网络进行拼接,还是需要再回到原文看一下细节.这里 ...
- Feature Pyramid Networks for Object Detection
Feature Pyramid Networks for Object Detection 特征金字塔网络用于目标检测 论文地址:https://arxiv.org/pdf/1612.03144.pd ...
- 论文阅读笔记三十三:Feature Pyramid Networks for Object Detection(FPN CVPR 2017)
论文源址:https://arxiv.org/abs/1612.03144 代码:https://github.com/jwyang/fpn.pytorch 摘要 特征金字塔是用于不同尺寸目标检测中的 ...
- 论文阅读 | FPN:Feature Pyramid Networks for Object Detection
论文地址:https://arxiv.org/pdf/1612.03144v2.pdf 代码地址:https://github.com/unsky/FPN 概述 FPN是FAIR发表在CVPR 201 ...
- FPN-Feature Pyramid Networks for Object Detection
FPN-Feature Pyramid Networks for Object Detection 标签(空格分隔): 深度学习 目标检测 这次学习的论文是FPN,是关于解决多尺度问题的一篇论文.记录 ...
- Parallel Feature Pyramid Network for Object Detection
Parallel Feature Pyramid Network for Object Detection ECCV2018 总结: 文章借鉴了SPP的思想并通过MSCA(multi-scale co ...
- FPN(feature pyramid networks)
多尺度的object detection算法:FPN(feature pyramid networks). 原来多数的object detection算法都是只采用顶层特征做预测,但我们知道低层的特征 ...
- Paper Reading: Relation Networks for Object Detection
Relation Networks for Object Detection笔记 写在前面:关于这篇论文的背景知识,请参考我前面的两篇随笔(<关于目标检测>和<关于注意力机制> ...
随机推荐
- max-width和width的区别
width为宽度 max-width为最大宽度 如果设置了width,那宽度就定死了,不能动态的改变,显得僵硬 而设置了max-width,实际宽度可以在0~max-width之间,当元素内部宽度不足 ...
- 国外线下技术俱乐部建设(1) - Belgrade Python技术俱乐部2019-01-25活动感悟
这是<国外线下技术俱乐部建设>系列文章之一. 虽然之前接触过Belgrade的.NET技术俱乐部,但是它最近活动要春节后了. 出于观摩别人是怎么搞线下社区的心态,还有自己也有在用Pyt ...
- 关于Xcode9.0版本模拟器Reset重置操作变更
- GIT的使用(Gitlab上传本地仓库代码,Webstorm修改更新)
准备:GIT的安装,Gitlab账户登陆,webstorm的安装 1.首先,你得先会在Gitlab中创建一个团体,在团体中创建一个项目,先建组,再建项目,网上哪里都有教程,随便找了个网址: https ...
- 为archlinux选择国内镜像
pacman-mirrors --country China && pacman -Syyu
- qt 打包发布 获取dll
发布前,获取所有qt dll包命令 生成的程序运行正常之后,找到项目的生成目录,比如 项目源码路径: C:\QtPros\hellomw\它的项目生成目录是C:\QtPros\build-hellom ...
- 爬虫系列----scrapy爬取网页初始
一 基本流程 创建工程,工程名称为(cmd):firstblood: scrapy startproject firstblood 进入工程目录中(cmd):cd :./firstblood 创建爬虫 ...
- 【Teradata SQL】FALLBACK表改为NO FALLBACK表
FALLBACK表在数据库中会留存双份数据,增加了数据可用性,但浪费了存储空间.变更表属性语句如下: alter table tab_fallback ,no fallback;
- 【Python 08】汇率兑换2.0-1(字符串索引)
1.案例描述 设计一个汇率换算程序,其功能是将人民币转换为美元,或者美元转换为人民币. 增加功能:根据输入判断是人民币还是美元,进行相应的转换计算. 2.案例分析 3.字符串 两个双引号或单引号括起 ...
- Linux系统中常见的目录名称以及相应内容
目录名称 应放置文件的内容 /boot 开机所需文件——内核.开机菜单以及所需配置文件等等 /dev 以文件形式存放任何设备与接口 /etc 配置文件 /home 用户家目录 /bin 存放单用户模式 ...