Feature Space Augmentation for Long-Tailed Data
Peng Chu Work was done at GE Research.11 Xiao Bian* 22 Shaopeng Liu 33 Haibin Ling 4411
Abstract
Real-world data often follow a long-tailed distribution as the frequency of each class is typically different. For example, a dataset can have a large number of under-represented classes and a few classes with more than sufficient data. However, a model to represent the dataset is usually expected to have reasonably homogeneous performances across classes. Introducing class-balanced loss and advanced methods on data re-sampling and augmentation are among the best practices to alleviate the data imbalance problem. However, the other part of the problem about the under-represented classes will have to rely on additional knowledge to recover the missing information.
中文速览
长尾分布（long-tailed distribution）是现实世界数据集中的普遍难题——少数类别样本充足，大量类别严重缺乏数据，导致模型在稀有类上表现很差。现有的重采样和平衡损失函数方法能在一定程度上缓解类别不平衡，但当尾部类别样本实在太少、无法覆盖其真实分布时，这些方法便力不从心。为此，作者提出利用类激活图（Class Activation Map，CAM）将每个类别的深层特征分解为"类别特有成分"和"类别通用成分"，再在训练过程中动态地将尾部类别的特有特征与头部类别（样本充足的类别）的通用特征进行融合，从而在特征空间中为尾部类别生成新的虚拟样本。在 iNaturalist、ImageNet-LT、Places-LT 及长尾版 CIFAR 等多个基准上，该方法均达到了当时的最优水平，证明借用头部类别学到的通用视觉知识来"脑补"尾部类别的分布是一条切实可行的路径，对提升现实场景中细粒度、长尾识别系统的性能具有重要意义。
原文 arXiv:2008.03673;中英对照 + 大白话阅读 https://aha.fim.ai/paper/2008.03673v1