[paper reading] rich feature hierarchies for accurate object detection and semantic segmentation

[ Address of thesis ][ Code ][CVPR 14]

Abstract

In the past few years , In a typical PASCAL VOC The target detection performance measured on the data set has stabilized . The best way to perform is a complex composite system , Usually, multiple low-level image features are combined with high-level background . In this paper , We propose a simple 、 Scalable detection algorithm , be relative to VOC 2012 The best result of , Average accuracy of the algorithm (mAP) Improved 30% above – Reached 53.3% Of mAP. Our approach combines two key insights ：(1) We can combine large capacity Convolutional Neural Networks (CNN) Apply to bottom-up regional recommendations (Region Proposal), In order to locate and segment the object ;(2) When the labeled training data is insufficient , Conduct supervised pre training for an auxiliary task , Then fine tune specific areas , Can produce significant performance improvements . Because we combine regional recommendations with CNN Combine , We call our approach R-CNN. We also put forward some experiments , These experiments provide insight into online learning content , It reveals a rich hierarchical structure of image features .

Method

This article is famous R-CNN, Among them R Express Region Proposal,. According to the current classification , This article is a two-stage Anchor Based Method , The process is as follows ：

• Step1: Region Proposal, That's the candidate area , It can be regarded as a box , The box contains items in the category we are interested in . In this paper, we use Selective Search Algorithm , Each image can get 2k～3k Candidate box .
• Step2: Because the box may contain objects we are interested in , So now it actually becomes a classification problem , Just judge the category of this box . therefore , Put the image in the box resize become CNN Acceptable input size ( Such as 224x224), And send it to CNN Feature extraction .
• Step3: Send the extracted features into SVM To classify , Judge its category . Be careful , because SVM Just a two classifier , So how many possible classes there are is to send a single feature into how many different SVM.

Pros

The highlights of this article are as follows ：

• Use deep learning model to extract features , Compared with the traditional manual features, the performance has achieved a leap .
• Introduced finetune thought , Through the image classification task ILSVRC The pre training on the data set gets rich representation , Then in the relatively small target detection data set VOC On finetune.

Cons

The deficiencies of this paper are as follows ：

• slow . Time is mainly spent on Selective Search Generate candidate boxes , as well as SVM Classification , And the feature is to be stored on the disk ( Take up space ). Therefore, the follow-up came one after another Fast R-CNN And Faster R-CNN.