Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

ML Tasks Image Detection SSD #154

Open
junxnone opened this issue Nov 21, 2018 · 1 comment
Open

ML Tasks Image Detection SSD #154

junxnone opened this issue Nov 21, 2018 · 1 comment

Comments

@junxnone
Copy link
Owner

junxnone commented Nov 21, 2018

SSD Single Shot MultiBox Detector

Reference

Brief

采用 VGG16 的基础网络结构,使用前面的前 5 层,然后利用 astrous 算法将 fc6 和 fc7 层转化成两个卷积层。再格外增加了 3 个卷积层,和一个 average pool层。不同层次的 feature map 分别用于 default box 的偏移以及不同类别得分的预测,最后通过 nms得到最终的检测结果。

SSD is an unified framework for object detection with a single network.

  • SSD提取了不同尺度的特征图来做检测,大尺度特征图(较靠前的特征图)可以用来检测小物体,而小尺度特征图(较靠后的特征图)用来检测大物体;
  • SSD采用了不同尺度和长宽比的先验框(Prior boxes, Default boxes,在Faster R-CNN中叫做锚,Anchors)
  • 原来VGG16中的fc6/7是全连接层,这里变成卷积层
  • MobileNetV2 paper 中提出 SSDLite

use_depthwise options for both box_predictor and feature_extractor

image

image

训练过程

image

样本选取

正样本 / 负样本
根据Feature map 确定Default Box 匹配ground truth 选择IOU 最大和 IOU>0.5 的作为候选正样本集, 其他的作为候选负样本集。
将每一个GT上对应prior boxes的分类loss 进行排序,选取loss 较低的样本作为正负样本。

Term

  • Feature map
  • Default box: 是每个小格上对应的一系列box。
  • Prior box : 是指实际中选取的default box。
  • jaccard overlap: 交并比

image

  • aspect ratio:横纵比
  • scale:大小
@junxnone
Copy link
Owner Author

junxnone commented Apr 2, 2019

#213 #85

@junxnone junxnone changed the title SSD - Single Shot MultiBox Detector ML Tasks Image Detection SSD Dec 24, 2022
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

1 participant