标签:第四课 5.5 AI 检测 目标 MMDetection 3.4 OpenMMLab CNN
OpenMMLab AI实战营 第四课笔记
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230202115351337-523975146.jpg)
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230203174116047-204135633.png)
目录
目标检测与MMDetection
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206210407131-527475089.png)
1.什么是目标检测
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206210516138-1877323993.png)
1.1 目标检测的应用
1.1.1 目标检测 in 人脸识别
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206210527670-1682427341.png)
1.1.2 目标检测 in 智慧城市
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206210813719-1749812769.png)
1.1.3 目标检测 in 自动驾驶
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206210828953-1083398390.png)
1.1.4 目标检测 in 下游视觉任务
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206210847589-1996266926.png)
1.2 目标检测 vs 图像分类
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206211056348-2007180370.png)
1.2.1 滑动窗口Sliding Window
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206211209810-1868917746.png)
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206211226957-105684303.png)
1.2.2 滑窗的效率问题
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206211244721-1889085197.png)
1.2.2.1 改进思路1:区域提议
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206211524474-639407115.png)
1.2.2.2 改进思路2:分析滑动窗口中的重复计算
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206211546050-1687404268.png)
1.2.2.3 消除滑窗中的重复计算
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206211602668-1123869895.png)
1.2.2.4 在特征图上进行密集预测
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206211617691-610011844.png)
1.3 目标检测的基本范式
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206211636506-1026833601.png)
1.4 目标检测技术的演进
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206211647857-1794416008.png)
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206211707645-229704527.png)
2.基础知识
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206211835850-1269939739.png)
2.1 框,边界框(Bounding Box)
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206211858505-2011491708.png)
2.2 框相关的概念
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206211914762-1639732460.png)
2.3 交并比(Intersection Over Union)
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206211935161-1954662072.png)
2.4 置信度
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206212138998-1083588074.png)
2.5 非极大值抑制
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206212147441-1878859583.png)
2.6 边界框回归
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206212323316-203820169.png)
2.7 边界框编码
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206212331739-519656262.png)
3.两阶段目标检测算法
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206212350244-364431499.png)
3.1 两阶段算法概述
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206212437485-1531240486.png)
3.2 Region-based CNN(2013)
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206212451312-988148960.png)
3.2.1 R-CNN的训练
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206212507421-1485682306.png)
3.2.2 R-CNN相比于传统方法的提升
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206212523756-2111881941.png)
3.2.3 R-CNN的问题
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206212534070-1745540352.png)
3.3 Fast R-CNN(2014)
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206212547732-1774016402.png)
3.3.1 Rol Pooling
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206212601223-1248140045.png)
3.3.2 Rol Align
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206212614475-804638504.png)
3.4 Fast R-CNN
3.4.1 Fast R-CNN的训练
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206212800849-1065592564.png)
3.4.2 Fast R-CNN的速度提升
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206212814441-355448479.png)
3.4.3 Fast R-CNN的精度提升
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206212921170-1914981141.png)
3.4.4 Fast R-CNN的速度瓶颈
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206212933475-1580228511.png)
3.4.5 降低区域提议的计算成本
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206212947012-1411831840.png)
3.4.6 朴素方法的局限
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213000594-307524314.png)
3.4.7 锚框Anchor
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213012778-186861234.png)
3.5 Faster R-CNN(2015)
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213023674-721404125.png)
3.5.1 Faster R-CNN的训练
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213032252-2005290878.png)
3.6 两阶段方法的发展与演进(2013~2017)
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213043072-1594543778.png)
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213109494-1077806714.png)
4.多尺度检测技术
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213135381-91667023.png)
4.1 多尺度检测必要性
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213242432-792144139.png)
4.2 图像金字塔Image Pyramid
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213256512-1113158574.png)
4.3 层次化特征
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213305848-1846800560.png)
4.4 特征金字塔网络Feature Pyramid Network(2016 )
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213333856-1771236802.png)
4.5 在Faster R-CNN模型中使用FPN
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213342507-1593747505.png)
5.单阶段目标检测算法
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213353221-1907162432.png)
5.1 回顾两阶段算法
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213418804-488038408.png)
5.2 单阶段算法
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213430392-544765650.png)
5.3 单阶段检测算法概述
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213447868-749508566.png)
5.4 YOLO:You Only Look Once(2015)
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213503092-1786874179.png)
5.4.1 YOLO的分类和回归目标
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213518389-1990025947.png)
5.4.2 YOLO的损失函数
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213530435-1461775878.png)
5.4.3 YOLO的优点和缺点
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213551177-1418812474.png)
5.5 SSD:Single Shot MultiBox Detector(2016)
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213608787-70940309.png)
5.5.1 SSD的损失函数
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213625634-425704945.png)
5.5.2 正负样本不均衡问题
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213642375-2002325421.png)
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213659613-1387074125.png)
5.5.3 解决样本不均衡问题
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213711837-677711603.png)
5.5.4 困难负样本Hard Negative
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213759640-1817489126.png)
5.5.5 不同负样本对损失函数的贡献
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213814886-692296099.png)
5.5.6 降低简单负样本的损失
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213826424-1781553116.png)
5.5.6 Focal Loss
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213840442-882656462.png)
5.6 RetinaNet(2017)
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213854385-1689301347.png)
5.6.1 RetinaNet的性能
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213909581-1265990946.png)
5.7 YOLOv3(2018)
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213921401-1815904438.png)
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213929241-1009053484.png)
6.无锚框目标检测算法
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213939631-1623926172.png)
6.1 锚框 vs 无锚框
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206213954377-293558567.png)
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214116471-1336320247.png)
6.2 FCOS,Fully Convolutinal One-Stage(2019)
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214129278-1014344688.png)
6.2.1 FCOS的多尺度匹配
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214141902-2086153085.png)
6.2.2 FCOS的预测目标
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214155653-1398052445.png)
6.2.3 中心度Center-ness
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214203659-937647413.png)
6.2.4 FCOS的损失函数
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214213954-891480305.png)
6.3 CenterNet(2019)
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214226408-1036357754.png)
6.3.1 CenterNet的主要流程
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214237419-247415458.png)
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214318981-1591334143.png)
7.1 DETR(2020)
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214331847-1024317032.png)
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214339274-533680411.png)
8.目标检测模型的评估方法Evaluaion
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214350397-512336548.png)
8.1 检测结果的正确/错误类型
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214455382-371669508.png)
8.2 准确率Rrecision与召回率Recall
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214526352-141773567.png)
8.3 准确率与召回率的平衡
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214538558-500025700.png)
8.4 PR曲线与AP值
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214605994-188522972.png)
8.5 完整数据集上的例子
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214630992-784474975.png)
8.6 PR曲线的起伏
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214640516-1374170178.png)
8.7 Mean AP
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214648257-1773129659.png)
8.8 总结
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214657522-968939591.png)
9.MMDetection
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214711979-254153639.png)
9.1 目标检测工具包MMDetection
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214722703-1761140122.png)
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214743989-96723981.png)
9.2 广泛应用
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214752549-71698701.png)
9.3 MMDetection可以做什么
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214826904-2132595040.png)
9.4 MMDetection环境搭建
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214835408-1329556292.png)
9.5 OpenMMLab项目中的重要概念——配置文件
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214844737-850603366.png)
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214856690-646029280.png)
9.6 MMDetection代码库结构
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214906659-770122354.png)
9.7 配置文件的运作方式
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214917714-755977259.png)
9.8 两阶段检测器的构成
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214928047-714359842.png)
9.9 单阶段检测器的构成
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206214939945-1753949082.png)
9.10 RetinaNet模型配置-主干网络
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206223753410-52657539.png)
9.10.1 RetinaNet模型配置-颈部
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206223802260-459511305.png)
9.10.2 RetinaNet模型配置-bbox head1
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206223815788-1344475004.png)
9.10.3 RetinaNet模型配置-bbox head2
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206223941494-1177564853.png)
9.11 COCO数据集介绍
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206223954348-50265259.png)
9.11.1 COCO数据集格式
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206224008007-1716801155.png)
9.11.2 COCO数据集的标注格式
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206224019673-698556491.png)
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206224050149-1412799694.png)
9.11.3 BBOX标注格式
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206224104006-741052985.png)
9.11.4 标注、类别、图像id的对应关系
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206224118046-772553225.png)
9.12 在MMDetection中配置COCO数据集
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206224126297-468961756.png)
9.12.1 MMDetection中的自定义数据集格式
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206224137095-1188929988.png)
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206224146248-454506414.png)
9.12.2 数据处理流水线
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206224201156-1579995795.png)
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206224213499-1489206317.png)
9.13 MMDetection中的常用训练策略
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206224226359-1051415694.png)
9.14 训练自己的检测模型
![](/i/l/?n=23&i=blog/1571518/202302/1571518-20230206224234059-967296233.png)
标签:第四课,
5.5,
AI,
检测,
目标,
MMDetection,
3.4,
OpenMMLab,
CNN
From: https://www.cnblogs.com/isLinXu/p/17096949.html