内容简介:This program shows how Bounding-Box-Regression works in a visual form.I am going to modify the above program in my free time, add this Bounding Box Regression along with ROI Align layer(
Bounding-Box-Regression-GUI
This program shows how Bounding-Box-Regression works in a visual form.
Updating soon for Digit-Recognition-CNN-and-ANN-using-Mnist-with-GUI
https://github.com/timmmGZ/Digit-Recognition-CNN-and-ANN-using-Mnist-with-GUI
I am going to modify the above program in my free time, add this Bounding Box Regression along with ROI Align layer( https://github.com/timmmGZ/ROIAlign-Bounding-Box-ROI-Align-of-Mask-RCNN-GUI ) in it, make it become a MNIST object detection.
First of all, Let's see how it works
Predefine
Download the MNIST digit 60000 train set and 10000 test set here:
https://drive.google.com/open?id=1VwABcxX0DaQakPpHbMaRQJHlJf3mVONf-
Put both files to ../dataset/, and then go to "tool" package.
-
Run the "CreatePictureForObjectDetectionFromMNIST.java", it will create random pictures base on MNIST datasets in ../dataset/pictures.
-
After step 2, run the "CreateForeOrBackgroundSample.java", it will create random foreground(in ../dataset/foreground) and background(in ../dataset/background) datasets based on output pictures of step 2, you could see it more clearly in ../dataset/groundTruthExamples.
as below:
-
Both step 2 and 3 will create Label-files in ../dataset/standardOutput, you could see the column names in first line of each Label-files
-
Make sure you have big enough RAM if you want to store more datasets in RAM, watch below picture:
Start
Run the MainFrame.java, if you don't want to train the model, click "Menu" then "Read Weight", that is my trained weights around 94% accuracy on both train and test set(70000 datasets in total, I define it is true prediction if Predicted-Bounding-Box has higher IOU with Ground-Truth-Bounding-Box than it has with input Bounding-Box).
Warning
Actually the number of predicted boxes should = the number of classes(e.g. one Bounding-Box can have both Apple and Bird inside it), but this program is just a Digit-Detection, for convenience and higher FPS in real-time detection, I make it have only one predicted box, see the advantage of having normal number of predicted boxes as below picture:
Bounding-Box-Regression is used after NMS(Non Maximum Suppression)
For each picture, we will get so many Bounding-Box(Region-Propasal), NMS is used for filtering out the best Bounding-Boxes, below gif shows what will look like if we only use ROI-Align-layer(not necessary but better use it) and NMS but not use Bounding-Box-Regression:
Predefine about ROI-Align-layer: sample size=1, output size=7, feature maps=16( https://github.com/timmmGZ/ROIAlign-Bounding-Box-ROI-Align-of-Mask-RCNN-GUI ) Let's cut one picture from the gif, when the object is small like below picture, and the number of objects is big, obviously we need to do Bounding-Box-Regression, or it will be a mess.
以上就是本文的全部内容,希望本文的内容对大家的学习或者工作能带来一定的帮助,也希望大家多多支持 码农网
猜你喜欢:本站部分资源来源于网络,本站转载出于传递更多信息之目的,版权归原作者或者来源机构所有,如转载稿涉及版权问题,请联系我们。
HTML 5实战
陶国荣 / 机械工业出版社 / 2011-11 / 59.00元
陶国荣编著的《HTML5实战》是一本系统而全面的HTML 5教程,根据HTML 5标准的最新草案,系统地对HTML 5的所有重要知识点进行了全面的讲解。在写作方式上,本书以一种开创性的方式使理论与实践达到极好的平衡,不仅对理论知识进行了清晰而透彻的阐述,而且根据读者理解这些知识的需要,精心设计了106个完整(每个案例分为功能描述、实现代码、效果展示和代码分析4个部分)的实战案例,旨在帮助读者通过实......一起来看看 《HTML 5实战》 这本书的介绍吧!