当前位置:网站首页>目标检测中的损失函数与正负样本分配:RetinaNet与Focal loss
目标检测中的损失函数与正负样本分配:RetinaNet与Focal loss
2022-07-07 00:26:00 【cartes1us】
RetinaNet
在目标检测领域,单阶段算法精度第一次超过双阶段,就是RetinaNet。
网络结构:
作者设计的网络结构没有太大创新,文中是这样说的:
The design of our RetinaNet detector shares many similarities with
previous dense detectors, in particular the concept of ‘anchors’
introduced by RPN [3] and use of features pyramids as in SSD [9] and
FPN [4]. We emphasize that our simple detector achieves top results
not based on innovations in network design but due to our novel loss.
检测头是分类与BBox回归解耦的,并且是基于锚框的,经过FPN后输出五层不同尺度的特征图,每层分别对应32~512尺度的锚框,并且每层根据scale和ratios的不同组合有9种锚框,最终整个网络的锚框尺寸是32 ~ 813之间。使用网络预测的相对锚框的偏移量来计算BBox的方法与Faster R-CNN相同。下图是霹雳吧啦的图。
论文中的结构图如下,只示意了由FPN引出的三种尺度的特征图,W,H,K,A分别代表特征图宽,高,分类数量(不包含背景类),锚框数量(9)。
正负样本匹配
正样本:预测的BBox 与gt IoU>=0.5,
负样本:预测的BBox 与gt IoU<0.4,
其他样本舍弃
前景,背景数量不平衡问题
CE损失的变种
本作的最大创新:Focal loss,改写经典的交叉熵损失,应用在class subnet分支,使易分类样本的损失的权重极大降低,形式很优美。论文中 γ \gamma γ推荐取2,若 γ \gamma γ取0,则FL就退化为了CE。
损失:
第一项分类损失中是计算所有样本(包括正负)的Focal loss,然后除去正样本数量 N p o s N_{pos} Npos。BBox回归损失是Fast R-CNN中提出的smooth L1 loss。
未完待续
边栏推荐
猜你喜欢
判断文件是否为DICOM文件
Pinduoduo product details interface, pinduoduo product basic information, pinduoduo product attribute interface
bat 批示处理详解
C nullable type
How does mapbox switch markup languages?
[reading of the paper] a multi branch hybrid transformer network for channel terminal cell segmentation
[论文阅读] Semi-supervised Left Atrium Segmentation with Mutual Consistency Training
基于NCF的多模块协同实例
SAP ABAP BDC (batch data communication) -018
AI人脸编辑让Lena微笑
随机推荐
Hcip seventh operation
Différenciation et introduction des services groupés, distribués et microservices
R language [logic control] [mathematical operation]
4. 对象映射 - Mapping.Mapster
The 2022 China low / no code Market Research and model selection evaluation report was released
"Multimodal" concept
拼多多商品详情接口、拼多多商品基本信息、拼多多商品属性接口
Message queue: how to handle repeated messages?
Getting started with DES encryption
Message queuing: how to ensure that messages are not lost
Reading the paper [sensor enlarged egocentric video captioning with dynamic modal attention]
消息队列:消息积压如何处理?
Message queue: how to deal with message backlog?
Flink SQL 实现读写redis,并动态生成Hset key
bat 批示处理详解
Differences and introduction of cluster, distributed and microservice
What is message queuing?
Jhok-zbl1 leakage relay
Leakage relay jelr-250fg
TCC of distributed transaction solutions