OD-Paper【1】：Rich feature hierarchies for accurate object detection and semantic segmentation

---

Preface

RCNN（Regions with CNN features） Yes, it will CNN This is a milestone in the application of the method to target detection , With the help of CNN Good feature extraction and classification performance , adopt RegionProposal Method to realize the transformation of target detection problem .

This paper mainly interprets the paper , And explain about RCNN Part of the frame

---

1. Abstract & Introduction

1.1. Abstract

The method of this paper combines two key factors ：

• Use a high-capacity convolutional neural network from bottom to top on the candidate region （CNNs）（ Network capacity ： The ability of artificial neural network to shape complex functions ）, Used to locate and segment objects .
• When the labeled training data is insufficient , First, carry out supervised pre training for auxiliary tasks, and the good model , Plus fine tuning on a specific domain , Can produce significant performance improvement .

In this paper, candidate regions （region proposals） And CNNs combination , Therefore, this method is called R-CNN（Regions with CNN features）： have CNN Candidate areas for features .

1.2. Introduction

This paper focuses on two issues ： Use deep network to locate targets and train high-capacity network models on small-scale labeled data sets .

Different from image classification , Detection requires locating many objects in an image . One way is to think of location as a regression problem . Another alternative is to build a sliding window detector , Use... In this way CNNs At least there are 20 It's been years . Usually used for some limited object categories , Like face , Pedestrians, etc . In order to maintain high spatial resolution , these CNNs Only two convolution layers and two pooling layers are used .

contrary , This paper is through the use of “ Use the identification of candidate areas ” Methods , It's solved CNN The problem of positioning （ This has achieved success in target detection and semantic segmentation ）. When testing , The method of this paper is input image , Generated close 2000 Category independent candidate areas , utilize CNN Extract a fixed length feature vector from each candidate region , then With the help of linearity specific to specific categories SVM Classify each candidate area . We don't consider the size of the area , The method of radiograph deformation is used to generate a fixed length image for each candidate region CNN Input eigenvector （ That is to put candidate areas of different sizes into the same size ）. The following figure shows an overview of our method and highlights some experimental results . Because our system combines candidate regions and CNNs, So it's called R-CNN（regions with CNN features）： have CNN Candidate areas for features .

---

2. Object detection with R-CNN

2.1. Algorithm process

RCNN The algorithm flow can be divided into 4 A step ：

• One image generation 1000～2000 Candidate areas （ Use Selective Search Method ）
• For each candidate area , Use deep networks to extract features
• Features are fed into each type of SVM classifier , Judge whether it belongs to this category
• Use regression to fine tune candidate box position

2.2. Module design

2.2.1. Region proposals

utilize Selective Search The algorithm obtains some original regions by image segmentation , Then use some merge strategies to merge these regions , Get a hierarchical regional structure , And these structures contain objects that might be needed .

2.2.2. Feature extraction

For each candidate area , Use deep networks to extract features . take 2000 Zoom candidate areas to 227 x 227 pixel , Then input the candidate area into the pre trained AlexNet CNN Internet access 4096 The feature of dimension is obtained 2000 x 4096 dimension matrix .

there CNN Image classification network , But remove the last full connection layer , Get the output 4096 Dimension vector .

2.2.3. SVM classifier

take 2000 x 4096 dimension Characteristics and 20 individual SVM The weight matrix 4096 x 20 Multiply , get 2000 x 20 dimension The matrix indicates that each suggestion box is the score of a certain goal . Respectively for the above 2000 x 20 dimension Each column in the matrix, that is, each class Non maximum suppression Eliminate overlapping suggestion boxes , Get this column, that is, some suggestion boxes with the highest scores in this category .

Be careful ： To facilitate observation , The vector in the picture is transposed

It can be understood as , Each column vector represents the probability that the category box in the picture is the class corresponding to the column

• The matrix on the left ： Each line is where each feature box passes CNN The feature vector obtained after classifying the network
• The matrix in the middle ： namely SVM A weight matrix , Each column corresponds to the weight vector of a category
• The matrix on the right ： Suppose the first column of the matrix in the middle represents a cat , The second column represents dogs , Then the first row of the matrix on the left is multiplied by the first column of the intermediate matrix , Get the elements of the first row and the first column of the matrix on the right , That is, the probability that the first candidate box is a cat
  • For each column of the matrix on the right （ That is, each type ）, Carry out non maximum suppression , To eliminate some overlapping suggestion boxes
  • For example, the first column of the matrix on the right represents the probability that all candidate boxes are cats

2.2.3.1. Non maximum suppression

IoU (Intersection over Union)

• The mathematical representation is ：$(A\cap B)/(A\cup B)$

• For the use of each category
  • Looking for the highest scoring target
  • Calculate the relationship between other goals and that goal IoU value
  • Delete all IoU Targets with values greater than a given threshold
    • The deleted targets here refer to those parts that overlap with the highest score much ,IoU Although big , But the score is not the highest , One with the highest score is enough
    • The purpose is to eliminate overlapping frames ,IoU The larger the value, the more overlapping areas

2.2.4. Regressor correction

Use regression to fine tune candidate box position , Yes NMS (Non-Maximum Suppression) After processing, the remaining suggestion boxes are further filtered . Then separately use 20 A regressor for the above 20 The suggestion boxes belonging to the category are regressed , Finally get the highest score of each category after correction bounding box.

After regression classifier , You'll get 4 Parameters ： The center point of the goal suggestion box x Offset 、y Offset , Scaling factor of bounding box Height 、 Width scaling factor

Pictured , Yellow frame $P$ A suggestion box Region Proposal, Green window $G$ Represents the actual box Ground Truth, The red window $\hat{G}$ Express Region Proposal The prediction window after regression , Linear regression problems that can be solved by the least square method .

---

3. R-CNN The problem is

• The test speed is slow
  Test a picture about 53s(CPU). use Selective Search The algorithm takes about 2s, There is a lot of overlap between candidate frames in an image , Feature extraction operation redundancy .
• Training is slow
  The process is extremely cumbersome
• Training needs a lot of space
  about SVM and bbox Back to training , We need to extract features from each target candidate box in each image , And write to disk . For very deep networks , Such as VGG16, from VOC07 From the training set 5k The features extracted from the image need hundreds of G Storage space

---

summary

Translation reference