Face Detection with the Faster R-CNN
来源:互联网 发布:研华数据采集卡 编辑:程序博客网 时间:2024/05/20 11:23
物体检测合集Object Detection blog
数据集合集CV Datasets on the web
正文
当我选用Faster RCNN做人脸检测时,已经有了一篇同名的论文了,详见Face Detection with the Faster R-CNN,同时其它博文也有用Faster RCNN训练自己数据集的教程,所以在此就不班门弄斧了。人脸检测的数据集我采用了FDDB,原因是该数据集小,而且之前用过。如果想要训练结果好一点,可以采用WIDER face,MegaFace这些大的数据集。
较其它博文的不同,我模仿Faster RCNN源代码手写了FDDB的训练过程(很多都是可重用的,需要写的不多)。我试过可以不改变solver就可以进行训练,如果出现loss为Nan的情况,可以适当减小lr试试。
图片来源于实验室,取自与某Q空间
训练结果比较差,在一些稍微暗一点或者人脸小一点的图片,就没法进行检测。本次练手的目的只是熟悉源码,对与训练结果没有做进一步优化,相关工作上面Face Detection with the Faster R-CNN也做过了,所以就不重复工作了。下次用SSD做个Face detection in video吧。
输入为data, im_info(height, width, scale)和gt_boxes。conv1-conv5为共享卷积层。rpn_cls_score(分类)和rpn_bbox_pred(回归)分别输出2 * 9(anchors)和4 * 9(anchors)。重点是”rpn-data”层anchor的操作。那么问题来了。
如何实现共享卷积层?
RPN和Fast R-CNN共享卷积层。RPN在features maps生成的候选窗口,然后经roi pooling后再检测,边框校准。
交替训练方法:
1. Stage 1 RPN, init from ImageNet model’
2. Stage 1 RPN, generate proposals
2. Stage 1 Fast R-CNN using RPN proposals, init from ImageNet model
3. Stage 2 RPN, init from stage 1 Fast R-CNN model
4. Stage 2 RPN, generate proposals
5. Stage 2 Fast R-CNN, init from stage 2 RPN R-CNN model
anchors的工作原理?
在conv feature map上做3*3的滑窗,每个窗口区域预测输入图像的3种尺度(128,256,512)和3种长宽比(1:1,1:2,2:1)的候选区。以40*60的feature map为例,总共有约20000(40*60*9)个anchor(候选区, region proposal)。针对人脸检测(大致为正方形),anchors可以调整长宽比。
Code
- Face Detection with the Faster R-CNN
- Face Detection with the Faster R-CNN (基于Faster RCNN的人脸检测)
- Face Detection with the Faster R-CNN(数据集标注对比研究报告 )
- Face Detection with the Faster R-CNN(数据集标注对比研究报告 )
- 【深度学习:目标检测】 Face Detection with the Faster R-CNN(数据集标注对比研究报告 )
- 论文解读Face Detection using Deep Learning: An Improved Faster R-CNN Approach
- 论文笔记:Faster R-CNN:Towards Real-Time Object Detection with Region Proposal Networks
- 论文笔记:Faster R-CNN:Towards Real-Time Object Detection with Region Proposal Networks
- Faster R-CNN:Towards Real-Time Object Detection with Region Proposal Networks
- Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks论文笔记
- 【Deep Learning】Review:Faster R-CNN- TowardsReal-Time Object Detection with Region Proposal Networks
- 论文笔记:Faster R-CNN:Towards Real-Time Object Detection with Region Proposal Networks
- 【论文笔记】Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks
- 论文笔记| Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks
- Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks----论文笔记
- Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks
- RCNN(四)Faster R-CNN:Towards Real-Time Object Detection with Region Proposal Networks
- 论文笔记:Faster R-CNN:Towards Real-Time Object Detection with Region Proposal Networks
- Spark & Zeppelin
- Linux Mini CRT堆栈管理的实现解析
- 分拆素数和<hdoj2098>
- java开发操作系统内核:让内核突破512字节的限制
- MIT6.824 课程学习初步
- Face Detection with the Faster R-CNN
- 安卓端发送短信
- 4-4 求自定类型元素的平均 (10分)
- IntelliJ IDEA中如何设置30分钟自动刷新svn
- 陈伟霆,赵丽颖,李易峰,青云志,分析。
- 美素数<hdoj4548>
- lintcode minimum-depth-of-binary-tree 二叉树最小深度
- MYSQL的wait_timeout与连接池相关
- listView分割线的相关设置