Target detection for long tail distribution -- balanced group softmax

Deal with long tailed target detection – Balanced Group Softmax

This time I will introduce an article CVPR2020 The article , Titled “Overcoming Classifier Imbalance for Long-tail Object Detection with Balanced Group Softmax”, It mainly solves the problem of long tail data distribution in target detection , The solution is also very simple .

Long tailed data

First , Long tailed data exist widely , Here we use COCO and LVIS Take two data sets as examples , As shown in the figure below ：

The abscissa is the index of the category , The ordinate is the number of samples in the corresponding category .

You can see , In these two data sets , There is an obvious long tail distribution .

Previous methods for dealing with long tail distribution

Here are some related works , Given by category ：

• Resampling based on data （data re-sampling）
  • Oversampling the tail data ：Borderline-smote: a new over-sampling method in im- balanced data sets learning
  • Delete the header data ：class imbalance, and cost sensitivity: why under-sampling beats over sampling
  • Sampling based on category balance ：Exploring the limits of weakly supervised pretraining.
• Cost sensitive learning （cost- sensitive learning）
  • Through to loss Adjustment , Give different weights to different categories

These methods are usually sensitive to hyperparameters , And poor performance when migrating to the detection framework （ The difference between classification task and detection task ）

Balanced Group Softmax

Here is the specific framework of the algorithm ：

As shown in the figure above , In the training phase , We will group the categories , Calculate separately in different groups Softmax, Then calculate the respective cross entropy error .

For grouping , The paper is given by 0,10,100,1000,+inf As a segmentation point

Here we need to add one for each group other Category , bring , When the target category is not in a group ,groundtruth Set to other.

The final error form is ：
$${\mathcal{L}}_k=-\sum _{n=0}^N\sum _{i\in {\mathcal{G}}_n}{y}_i^n\log (p_i^n)$$
among ,$N$ It's the number of groups ,${\mathcal{G}}_n$ It's No $n$ Category collection of groups ,$p_i^n$ Is the probability of model output ,$y_i^n$ Is the label .

Effect evaluation

Here is a precision table for comprehensive comparison

AP The subscript of corresponds to the index of the divided group , You can see , Precision in the tail , That is to say $A{P}_1$ and $AC{C}_1$ It has reached SOTA Performance of .