目标检测--Single-Shot Refinement Neural Network for Object Detection

Single-Shot Refinement Neural Network for Object Detection
https://github.com/sfzhang15/RefineDet

针对目标检测,本文可以看作将 Faster RCNN 和 SSD 融合起来。

1 Introduction
当前基于 CNN 网络的目标检测可以分为两大类:1) the two-stage approach,2)the one-stage approach
1) the two-stage approach
首先是候选区域的提取,然后是目标的分类和回归,这类方法的检测精度要好于the one-stage approach
2)the one-stage approach
这类方法是通过 regular and dense sampling over locations, scales and aspect ratios,速度快,精度稍微差些,主要原因是 the class imbalance problem,即正负样本比例严重失调

本文提出一个目标检测框架 RefineDet, to inherit the merits of the two approaches (i.e., one-stage and two-stage approaches) and overcome their shortcomings

Architecture of RefineDet
目标检测--Single-Shot Refinement Neural Network for Object Detection_第1张图片

网络主要包括两个相互关联的模型 Anchor Refinement Module (ARM)和 Object Detection Module(ODM),这两个模块通过 transfer connection block (TCB) 联系起来。

Anchor Refinement Module (ARM) 可以看作一个简化的 SSD,这里只做二分类,即目标的有无,去除一些无物体的候选区域,对位置和尺寸进行大致的调整,为后面的 ODM 提高一个好的初始化
ARM aims to remove negative anchors so as to reduce search space for the classifier and also coarsely adjust the locations and sizes of anchors
to provide better initialization for the subsequent regressor

Object Detection Module(ODM) 可以看作为一个 Fast RCNN, 多类别分类是和矩形框回归
ODM aims to regress accurate object locations and predict multi-class labels based on the refined anchors

Transfer Connection Block (TCB) 这个模块可以看作 FCN 中的 deconvolution layers, 将不同网络层的特征融合起来
目标检测--Single-Shot Refinement Neural Network for Object Detection_第2张图片

5 Experiments
目标检测--Single-Shot Refinement Neural Network for Object Detection_第3张图片
RefineDet320+ 、RefineDet512+ : multi-scale testing strategy

目标检测--Single-Shot Refinement Neural Network for Object Detection_第4张图片

目标检测--Single-Shot Refinement Neural Network for Object Detection_第5张图片

11

你可能感兴趣的:(目标检测)