blackfeather-wang / ISDA-for-Deep-Networks

An efficient implicit semantic augmentation method, complementary to existing non-semantic techniques.
582 stars 94 forks source link

long-tailed problem #20

Closed AaboutL closed 3 years ago

AaboutL commented 3 years ago

你好, 谢谢你的分享。我有一个问题,这个增强方法对于long-tailed problem是否work,或者说,如果很多类别中的样本仅仅只有极少的几张图片,这个方法能起作用吗?

blackfeather-wang commented 3 years ago

您好,

非常感谢您的关注,事实上long-tailed problem + ISDA也是我们正在关注的问题,我们刚刚完成了一项工作,发表在cvpr2021,已经挂到Arxiv上面: MetaSAug: Meta Semantic Augmentation for Long-Tailed Visual Recognition。

根据我们目前的结果,ISDA直接应用于long-tailed problem 是可以改善表现的,只不过其效果可能比较有限,主要原因是对于样本较少的类别协方差矩阵估计的不大好,因此是需要一定的改进的。