[target detection] generalized focal loss v1

The paper ： https://arxiv.org/pdf/2006.04388.pdf

Code ：https://github.com/open-mmlab/mmdetection/tree/master/configs/gfl

Source ：NIPS2020

emphasis ：

• A new method for determining the position of the bounding box is proposed generalize Modeling of distribution （ The clearer the boundary, the better the learning , The distribution will be sharp , The more fuzzy the boundary is, the worse the learning will be , Flat distribution ）

One 、 background

One-stage Target detector basically models target detection as a task of dense classification and location .

Classification tasks generally use Focal Loss To optimize , Positioning tasks are generally learning Dirac delta Distribution .

Such as FCOS A quantity for estimating the positioning quality is proposed in ：IoU score or centerness score, then NMS When sorting , Multiply the classification score by the box quality score .

Current One-stage The target detector usually introduces a separate prediction branch to quantify the positioning effect , The prediction effect of positioning is helpful for classification , So as to improve the detection performance .

This paper proposes three basic elements ：

• Quality estimation of detection frame （ Such as IoU score or FCOS Of centerness score）
• classification
• location

There are two main problems in the current implementation ：

1、 The classification score and frame quality estimation are inconsistent during training and testing

• Inconsistent usage ： Classification and quality estimation , In the training process is separate , But in the test process, it is multiplied together , As NMS score Sort by , There is a certain gap

• The objects are not the same ： With the help of Focal Loss Power , Classification and branching can make a small number of positive samples and a large number of negative samples train together , But the quality estimation of the box is actually only for positive sample training .

  about one-stage detector , do NMS When sorting , All samples will multiply the classification score by the box quality score , To sort , Therefore, there must be some negative samples with low scores whose quality prediction has no supervision signal in the training process , That is, the quality of a large number of negative samples is not measured . This will lead to a negative sample with a low classification score , Due to the prediction of a very high box quality score , As a result, it is predicted to be a positive sample .
  

2、bbox regression The expression of is not flexible （Dirac delta Inflexible distribution ）, There is no way to model complex scenes uncertainty

• In a complex scene , The representation of bounding box has strong uncertainty , The essence of the existing box regression is to model a very single Dirac distribution , Very inflexible . So the author hopes to use a general To model the representation of the bounding box . The problem is shown in the figure 3 Shown （ Like a skateboard blurred by water , And heavily sheltered elephants ）：

Two 、 Method

For the two existing problems ：

① Training and testing are inconsistent

② The modeling of frame position distribution is not universal

The author proposes the following solutions .

Solve problem one ： Build a classification-IoU joint representation

For the first inconsistency between training and testing , In order to ensure the consistency of training and testing , At the same time, both classification and frame quality prediction can be trained to all positive and negative samples , The author proposes to combine the expression of box with classification score .

Method ：

When the category of prediction is ground-truth When it comes to categories , Use position quality score As confidence , The position quality score in this paper is to use IoU Score to measure .

Problem solving II ： Directly regress an arbitrary distribution to model the representation of the box

Method ： Use softmax To achieve , It involves deriving from the integral form of Dirac distribution to the integral form of general distribution to express the box

thus , It eliminates the inconsistency between training and testing , And established as shown in the figure 2b Strong correlation between classification and positioning .

Besides , Negative samples can be used 0 quality scores To supervise .

Generalized Focal Loss The composition of the ：

• QFL：Quality Focal Loss, Joint expression of learning classification score and position score
• DFL：Distribution Focal Loss, Model the position of the box as a general distribution, Let the network quickly focus on the distribution of positions close to the target position

Generalized Focal Loss How it was put forward ：

① original FL：

Today's intensive forecasting tasks , Generally used Focal Loss To optimize the classification and branch , Can solve the prospect 、 Problems such as the imbalance of the number of backgrounds , The formula is as follows , But it can only support 0/1 Such discrete categories label.

**① Put forward QFL：Quality Focal Loss **

The standard one-hot The code is 1, Other positions are 0.

Use classification-IoU features , Be able to put the standard one-hot Coding softens , Make it more soft, The goal of learning $y\in [0,1]$, Rather than direct learning objectives “1”.

For this paper, joint representation ,label Turned into 0~1 Continuous values of .FL No longer applicable .

• y=0 when , Negative samples ,quality score by 0
• 0<y<=1 when , Indicates a positive sample , And position score label y It's using IoU score It means , be in 0~1 Between

In order to ensure QFL Yes Focal Loss The balance of difficult and easy samples 、 The ability of positive and negative samples , It can also support the supervision of continuous values , Need to be right FL Make some extensions .

• Cross entropy $-log(p_t)$ An extension of ：$-((1-y)log(1-\sigma )+ylog(\sigma ))$
• Modulation factor $(1-p_t{)}^{\gamma }$ An extension of ：$|y-\sigma {|}^{\beta }(\beta >=0)$

Quality Focal Loss（QFL） Ultimately for ：

• $\sigma =y$ yes QFL Global minimum solution
• chart 5a It shows the difference $\beta$ The effect of （y=0.5）
• $\Vert y-\sigma {\Vert }^{\beta }$ Is a modulation factor , When a sample quality When the estimation is inaccurate , The modulation factor will be very large , Let the network pay more attention to this difficult sample , When quality When the estimate of tends to be accurate , namely $\sigma$ → $y$ when , The modulation factor tends to 0, This sample pair loss The influence weight of will be reduced .$\beta$ Control the process of reduction , this paper $\beta =2$ The optimal .

② Put forward DFL: Distribution Focal Loss

The position learning in this paper takes the relative offset as the regression goal , And the previous articles are generally based on Dirac distribution $\delta (x-y)$ For guidance , Satisfy ${\int }_{-\infty }^{+\infty }\delta (x-y)dx=1$, We usually use the full connection layer to realize .

But this paper takes into account the diversity of real distribution , Choose to use a more general distribution to represent the location distribution .

The real distribution is usually not too far away from the location of the annotation , So another one is added Loss

• DFL It can make the network focus on the target faster $y$ Nearby values , Increase their probability
• The meaning is to optimize and label in the form of cross entropy y The probability of the closest left and right positions , So that the network can focus on the distribution of adjacent areas of the target location faster

QFL and DFL It can be expressed as GFL：

• Variable is $y_l$ and $y_r$
• The predicted distributions of the above two variables are ： $p_{y_l}$ and $p_{y_r}$, And $p_{y_l}+p_{y_r}=1$
• The final prediction is ：$\hat{y}=y_l{p}_{y_l}+y_r{p}_{y_r}$, And $y_l<=\hat{y}<=y_r$

Trained loss as follows ：

3、 ... and 、 effect