Shulou(Shulou.com)06/01 Report--

Big data in how to solve the problem of long tail distribution and decoupling category features and achieve spatial expansion, many novices are not very clear about this, in order to help you solve this problem, the following editor will explain for you in detail. People with this need can come and learn. I hope you can get something.

Real-world data usually follow a long-tailed distribution, which means that the number of each class is usually different. For example, there are a large number of samples in the header class of a dataset and a small number of samples in the tail class. However, we hope that the model can fairly represent the entire data set, rather than favor some classes with large samples. Among the various methods to solve the long tail problem, class balance loss, resampling and data augmentation are the more common methods. But for the tail class, we have to consider some other knowledge to make up for the lost information. In this paper, the author proposes a new method to solve the long tail problem, which extends the features of the head class to the tail class in the feature space. Specifically, the characteristics of each class are decoupled into class-specific and class-common features, and the class-specific features of the tail class and the class common features of the head class are fused, so as to realize the expansion of the feature space. This method has achieved good results on four datasets: iNaturalist, ImageNet-LT, Places-LT and long-tailed CIFAR set.

Brief introduction

The distribution of long tail exists widely in visual tasks. As shown in the figure, there is a problem of long tail distribution in classification and detection tasks.

The main problem caused by the long-tailed distribution is that during training, because the amount of data of the tail class is small and the statistical information is not rich enough, the model can not express the tail class very well. The existing methods mainly include data augmentation, downsampling, oversampling, and the construction of balance loss function. However, the performance of these methods is not ideal when the number of tail classes is very small. This is shown in the following figure.

This paper proposes to transfer the information of the header class to the tail class in the feature space. The specific methods are as follows.

The method of this paper

In this paper, firstly, the attention region is extracted by CAM method, and the class-specific features and class common features of each class are obtained. After that, the unique characteristics of the tail class and the common characteristics of the head class are fused.

First of all, let's introduce CAM (Class Activation Map).

M is the highlight we got. C is the category, XQuery y is the pixel position, k is the channel, w is the weight, and f is the eigenvector. The larger the M is, the more important the features at XMagi y are for the category c. Then we normalize M to 0-1 and give a threshold.

Through the following formula, we can get the class-specific feature (s for specific) and the class common feature (g for generic)

Where it represents the Hadamard product, and when x is greater than or equal to 0, sgn (x) = 1, and when less than 0, sgn (x) = 0.

Then let's take a look at the overall training process. The first step is to train all the data to get the sub-network and basic classifier for feature extraction, which can be used in the following steps. The second step is to enlarge the tail class according to the extracted feature network and classifier, as well as the previous CAM. As you can see, the second step is to go into a tail picture and a head picture, and select the head class which is close to the tail class and easy to be confused (according to the order of confidence).

The third step is fine-tuning. Note that the third step and the second step are carried out synchronously, collectively referred to as the second stage. The overall algorithm flow of the second stage is as follows.

Experiments and results

Dataset: Long-tailed CIFAR-10 and CIFAR-100, ImageNet-LT and Places-LT Dataset,iNaturalist 2017 and 2018.

Comparative experiment:

Ablation experiment

Result analysis

Is it helpful for you to read the above content? If you want to know more about the relevant knowledge or read more related articles, please follow the industry information channel, thank you for your support.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.