author | Heat lamp lamp

The main purpose of this article is correct FPN Big target performance damage , It is found that the reason mainly comes from FPN Inappropriate in back-propagation, Therefore, the author puts forward the auxiliary loss based on uncertainty and constructs a new FPN Structure to solve the above problems .

Paper title ：

You Should Look at All Objects

The conference ：

ECCV 2022

Thesis link ：

https://arxiv.org/abs/2207.07889

Code link ：

https://github.com/CharlesPikachu/YSLAO

Preface

FPN (Feature Pryramids Network) Integration into the backbone network can effectively improve the performance of the target detection model , It is one of the basic modules of target detection , There have been a lot of related work to optimize FPN Structure , such as PANet and Nas-fpn. At present, I think FPN There are two main advantages , One is FPN By integrating the multi-layer characteristics of the backbone network , Can be better characterized , The other is FPN By dealing with targets of different sizes at each different stage , Realized the idea of partition . Obviously, these methods should be able to improve the detection effect of all scale targets , But the truth is Small and medium goals AP Got a promotion , Big target AP It's down .

As shown in the figure above , The author in MMDetection and Detectron2 Two detection frameworks are used for experimental comparison FPN Changes of various performance indicators before and after , You can see the use of FPN after , whole AP There is a promotion , But the big goal (AP_l) Have decreased in varying degrees . How to solve this problem ？ The author found FPN The change brought about is not only the multi-level feature fusion and divide and conquer thought , also back-propagation path Changes , This will also directly affect the effect of the detection model .

There is a problem here , The two models compared by the author are ResNet-50-DC5 and ResNet-50-FPN, It doesn't feel very fair ,ResNet-50-FPN Should be and ResNet-50 Comparison is right , Here we add FPN Only a few points higher , commonly FPN The rising point will be more . there DC5 It should be to add hole convolution behind the backbone network , The output receptive field behind is larger, and the detection effect of large targets should be better .

The author proposes to introduce more auxiliary loss functions to expand back-propagation path So that the additional supervised signals can assist the learning of the corresponding backbone network layer . The key technique is to use uncertainty to balance a large number of loss functions . besides , The author designs a new FPN Network to modify back-propagation path. The proposed method can steadily improve on multiple detection methods 2 percentage , Include one-stage, two-stage, anchor-based, anchor-free The detector .

First, a brief review FPN, It mainly consists of three parts ：top-down, bottom-up and lateral connection, The main differences with the basic backbone network are shown in the figure above . The author's main findings are FPN-free The shallow layer of the detection frame is back-propogation No effective supervision signal was received at the time of , However FPN-based All backbone network features of the detection framework can get direct supervision signals . To prove this phenomenon , The author did a simple experiment , That is to FPN-free The characteristic layer of the detector adds some auxiliary supervision loss .

The experimental results are shown in the figure above , It can be seen that after adding auxiliary losses ,FPN-free and FPN-based There is no difference in the performance of the detector （FPN-Aux v.s. DC5-Aux）. So why FPN It will also inhibit the detection performance of large targets ？ as a result of FPN The lowest feature in the structure is mainly used to supervise the learning objectives , So I learned feature It has good detection ability only for small targets . meanwhile , The bottom features will spread backwards , The ability to detect large targets will continue to weaken with the spread , So the final result Feature Relatively weak detection ability for large targets .

Method

2.1 Ancillary loss

Firstly, an auxiliary loss with uncertainty is proposed ：

The loss function here includes the loss of classification and regression , among α  It's uncertainty , Is obtained through a prediction function , as follows ：

here (x) yes feature map, (w) and (b) It's a learnable parameter . What's not clear is , Why did you get it like this α It can represent uncertainty ？

2.2 Feature pyramid generation paradigm

The main purpose of the author is to let the supervision loss see every goal . There are two changes here , One is Feature Grouping, The other is Cascade Structure. Simply speaking ,Feature Grouping Through some operations, the interaction between multiple layers is more , Mainly through some operations to exchange different layers feature Of channel. Maybe several layers of network get a feature converter to transform feature, Then press channel grouping , Finally, cross layer exchange channel.Cascade structure It is the intermediate features obtained above that are converted again through several layers of Nonlinear Networks . The specific operation is as follows ：

experiment

▲ Auxiliary loss Ablation Study, The introduction of uncertainty has been significantly improved

▲ About feature pyramid generation paradigm Of Ablation Study

It can be seen that , Combine Feature Grouping and Cascade The structure can be improved obviously , and Cascade The more times, the better .

summary

The article raises a question , namely FPN It is harmful to the detection of large targets , And found that the main reason is back-propagation path in , So two strategies are proposed to solve this problem , One is to add multi-layer auxiliary losses , And use uncertainty to balance the relationship between multiple losses . Moreover, through modification PFN To change the structure of back-propagation path, Put forward Feature Grouping To make the Feature more conpactness, Fully integrate the characteristics of different layers , Let the features of each layer see all Objects, The experimental results have been significantly improved on many detectors . 

There are also some problems in the article that are not very clear ： 

1. Fig.1 Chinese comparative experiment , use ResNet-50-DC5 And ResNet-50-FPN Is it appropriate to compare ？DC5 It may be helpful for the detection of large targets , If so , The article is about FPN The conclusion that it is harmful to large target detection remains to be discussed .

2. The article is actually right FPN Some improvements to , But not with others FPN Compare the work of , such as PANet, Nas-fpn, The problem of big goals may lie in these methods somehow It's been solved .

Read more

# cast draft   through Avenue #

  Let your words be seen by more people  

How to make more high-quality content reach the reader group in a shorter path , How about reducing the cost of finding quality content for readers ？ The answer is ： People you don't know .

There are always people you don't know , Know what you want to know .PaperWeekly Maybe it could be a bridge , Push different backgrounds 、 Scholars and academic inspiration in different directions collide with each other , There are more possibilities . 

PaperWeekly Encourage university laboratories or individuals to , Share all kinds of quality content on our platform , It can be Interpretation of the latest paper , It can also be Analysis of academic hot spots 、 Scientific research experience or Competition experience explanation etc. . We have only one purpose , Let knowledge really flow .

  The basic requirements of the manuscript ：

• The article is really personal Original works , Not published in public channels , For example, articles published or to be published on other platforms , Please clearly mark  

• It is suggested that  markdown  Format writing , The pictures are sent as attachments , The picture should be clear , No copyright issues

• PaperWeekly Respect the right of authorship , And will be adopted for each original first manuscript , Provide Competitive remuneration in the industry , Specifically, according to the amount of reading and the quality of the article, the ladder system is used for settlement

  Contribution channel ：

• Send email ：[email protected] 

• Please note your immediate contact information （ WeChat ）, So that we can contact the author as soon as we choose the manuscript

• You can also directly add Xiaobian wechat （pwbot02） Quick contribution , remarks ： full name - contribute

△ Long press add PaperWeekly Small make up

Now? , stay 「 You know 」 We can also be found

Go to Zhihu home page and search 「PaperWeekly」

Click on 「 Focus on 」 Subscribe to our column

·