Target detection | target size prediction based on statistical adaptive linear regression

Computer Vision Institute column

author ：Edison_G

YOLOv2 and YOLOv3 It is a typical target detection algorithm based on deep learning , They use statistical adaptive exponential regression model to design the last layer of the network to predict the size of the target .

Long press scan QR code, follow us

One 、 Briefly

What we share today is that researchers have proposed a target size prediction method based on statistical adaptive linear regression .YOLOv2 and YOLOv3 It is a typical object detection algorithm based on deep learning , They use statistical adaptive exponential regression model to design the last layer of the network to predict the size of the object .

However , Because of the properties of exponential function , Exponential regression model can spread the derivative of loss function to all parameters in the network . Researchers propose a statistical adaptive linear regression layer to alleviate the gradient explosion problem of exponential regression model . The proposed statistical adaptive linear regression model is used at the last layer of the network to predict the size of the target estimated from the statistical data of the training data set . Researchers have newly designed YOLOv3tiny The Internet , It's in UFPR-ALPR Data set is better than YOLOv3 It has higher performance .

Two 、 background

This is mainly a Korean paper , It makes my reading process more difficult , So today I will briefly introduce some overall framework ideas , Interested students can further read the paper , Deepen understanding ！

Today we won't introduce the traditional detection network , Because we ” Institute of computer vision “ I shared too many dry goods and practices of target detection before , If you want to get more familiar with the entry level, please check the history sharing . See the following link for some sharing ：

  Previous recommendation  

• Target detection dry goods | Multi level feature reuse greatly improves the detection accuracy （ At the end of the paper is attached the thesis download ）
• SSD7-FFAM | Embedded friendly target detection network , For the safety of kindergarten children
• A new way of target detection | class-agnostic The detector is used for target detection （ A link to download the paper is attached ）
• dried food | Using hand-held camera image, real-time rice detection is carried out through convolution neural network （ Pay homage to Mr. Yuan ）
• CVPR 2021 | No need to label ？ See how self supervised learning framework can help target detection
• object detection | Rich feature orientation Refinement Network For target detection （ attach github Source code ）

3、 ... and 、 Frame analysis

Above, YOLOv2 and YOLOv3 The post-processing process .

The proposed method uses the estimable statistical data in the learning data set to predict the width and height of the target , This is related to YOLOv2 and YOLOv3 identical . The process of estimating the statistical value of the learning data set is as follows : According to the width and height of the target in the learning data set , Classify targets as K A cluster of , Then estimate the arithmetic mean of the width and height of the target in each cluster . 

The statistical adaptive linear regression model proposed by the researchers further estimates the standard deviation of the width and height of the target in each cluster . then , The mean and standard deviation of the width and height of the target predicted by the network are designed to follow the mean and standard deviation of the width and height of the target in the learning data set . Use the estimable statistical values in the learning data set to constrain the statistical values of the predicted values , It can make the network more stable in the learning stage , Improve detection performance .

The existing YOLOv2, As for YOLOv3 Modification of statistical adaptive exponential regression model for target size prediction in , The researchers redefined the loss function of the statistical adaptive linear regression model for learning the proposed target size prediction , The proposed loss function is as above .

Four 、 experiment

UFPR-ALPR dataset

The network architecture of newly designed YOLOv3 tiny for experiments

The comparison on UFPR-ALPR test dataset

 THE END 

Please contact the official account for authorization.

The learning group of computer vision research institute is waiting for you to join ！

The Institute of computer vision is mainly involved in the field of deep learning , Mainly devoted to face detection 、 Face recognition , Multi target detection 、 Target tracking 、 Image segmentation and other research directions . The Research Institute will continue to share the latest paper algorithm new framework , The difference of our reform this time is , We need to focus on ” Research “. After that, we will share the practice process for the corresponding fields , Let us really experience the real scene of getting rid of the theory , Develop the habit of hands-on programming and brain thinking ！

Sweep yards attention

Institute of computer vision

official account ID｜ComputerVisionGzq

Study Group ｜ Scan the code to get the join mode on the homepage

Source download |  reply “OSP” Get Downloads