目标检测论文阅读笔记:《ThunderNet: Towards Real-time Generic Object Detection on Mobile Devices》
分类:
文章
•
2024-02-01 11:27:52
1 概述
- 本文基于ShuffleNetV2提出了一个轻量目标检测网络,称为ThunderNet,主要有两个创新点:①CEM(Context Enhancement Module),与FPN类似,用于融合浅层特征和深层特征;②SAM(Spatial Attention Module),用于增强目标特征
- 轻量化体现在两方面:①将ShuffleNetV2中3x3卷积更换为5x5卷积,在增大感受野使特征包含更多空间信息的同时降低参数量;②移除conv5的同时在之前的stage中增加channel数,这在不增加参数量的情况下可使浅层特征增多。
2 网络框架
2.1 整体框架
![目标检测论文阅读笔记:《ThunderNet: Towards Real-time Generic Object Detection on Mobile Devices》 目标检测论文阅读笔记:《ThunderNet: Towards Real-time Generic Object Detection on Mobile Devices》](/default/index/img?u=aHR0cHM6Ly9waWFuc2hlbi5jb20vaW1hZ2VzLzU4OS9lZWYwMTE5MGJmNjI4OTA4ZTE0OTRiNDI0OWEzMDRjNS5wbmc=)
2.2 CEM模块
![目标检测论文阅读笔记:《ThunderNet: Towards Real-time Generic Object Detection on Mobile Devices》 目标检测论文阅读笔记:《ThunderNet: Towards Real-time Generic Object Detection on Mobile Devices》](/default/index/img?u=aHR0cHM6Ly9waWFuc2hlbi5jb20vaW1hZ2VzLzE0MS85ZGRlNjcyZjU4OGM2ZTM1OTUyZTc3OTdkMDg1NmU3ZC5wbmc=)
- 类似FPN,其意义在于融合多个尺度信息,使生成的特征包含更多的空间信息
2.3 SAM模块
![目标检测论文阅读笔记:《ThunderNet: Towards Real-time Generic Object Detection on Mobile Devices》 目标检测论文阅读笔记:《ThunderNet: Towards Real-time Generic Object Detection on Mobile Devices》](/default/index/img?u=aHR0cHM6Ly9waWFuc2hlbi5jb20vaW1hZ2VzLzIxMC8zNWE0ZWVkMDk2NmE4ZDE3NjgzMzY0ZGM3OTc1Y2QyMi5wbmc=)
- RPN用于分类前景背景,其中背景特征会被抑制,因此论文中提出将RPN特征作为监督信号用来抑制CEM特征中的背景部分,增强CEM特征中的前景部分。
- 该结构的另外一个优点是反向传播过程中这里也会产生梯度,也算是RPN结构参数更新时的监督信号
3 实验
![目标检测论文阅读笔记:《ThunderNet: Towards Real-time Generic Object Detection on Mobile Devices》 目标检测论文阅读笔记:《ThunderNet: Towards Real-time Generic Object Detection on Mobile Devices》](/default/index/img?u=aHR0cHM6Ly9waWFuc2hlbi5jb20vaW1hZ2VzLzI3LzMwMmVlMWFkYjQ1YTNkNGUwNDMxZTc0NThlZGU0YzdiLnBuZw==)
![目标检测论文阅读笔记:《ThunderNet: Towards Real-time Generic Object Detection on Mobile Devices》 目标检测论文阅读笔记:《ThunderNet: Towards Real-time Generic Object Detection on Mobile Devices》](/default/index/img?u=aHR0cHM6Ly9waWFuc2hlbi5jb20vaW1hZ2VzLzUyOC85YTMyY2ZiMjE0OTMyMjY1OTIzNmQyYjliNGQxOGRlOC5wbmc=)
![目标检测论文阅读笔记:《ThunderNet: Towards Real-time Generic Object Detection on Mobile Devices》 目标检测论文阅读笔记:《ThunderNet: Towards Real-time Generic Object Detection on Mobile Devices》](/default/index/img?u=aHR0cHM6Ly9waWFuc2hlbi5jb20vaW1hZ2VzLzU4NS9iNDU1ZDE4NDFiZTRhYjA2ZjhhZmIzNDI4ODQxYmJmMS5wbmc=)
- 消融实验
- SAM模块效果
![目标检测论文阅读笔记:《ThunderNet: Towards Real-time Generic Object Detection on Mobile Devices》 目标检测论文阅读笔记:《ThunderNet: Towards Real-time Generic Object Detection on Mobile Devices》](/default/index/img?u=aHR0cHM6Ly9waWFuc2hlbi5jb20vaW1hZ2VzLzM0NC8yYWRmNzA4YzQwMWZkMzRiYmIwZjEyNjJjZTNmOGI3OC5wbmc=)
4 总结
- CEM中增大特征感受野的思路不错,也可以使用空洞卷积等,可以借鉴
- 将RPN特征作为监督信号引入很新颖,值得学习