Skip to content

Instantly share code, notes, and snippets.

@YimianDai

YimianDai/Noun-Discrimination.md

Last active August 2, 2019 23:19
Show Gist options
  • Save YimianDai/ead926792d30bec7d8027cac5f80760b to your computer and use it in GitHub Desktop.
Save YimianDai/ead926792d30bec7d8027cac5f80760b to your computer and use it in GitHub Desktop.
Download ZIP
名词辨析
Raw
Noun-Discrimination.md

Localization vs Detection

Localization 虽说都要给出 Bounding Box，但是不同于 Detection 有专门的 Detection 数据集，Localization 用的还是 Classification 的数据集，也就是每幅图像中只有一个占据主要区域面积的物体，而 Detection 的数据集中一幅图像内可以有任意多个物体，也可以没有。

ICLR-2014-Overfeat 那篇论文提到了两者的区别。

Loss vs Metric

  1. Loss 要可微分，这样才可以做 SGD；Metric 则没有这个要求
  2. Loss 最后是变成一个 scalar (某种程度上对 Overall Accuracy 的反应)，在 Class Imbalance 的情况下，数量小的类别会 get less prioritised in an effort to achieve the best overall accuracy，这就是 Focal Loss 的动机；而在 Metric 里面，比如 mAP，如果一个类很低，就会特别拉低。

Reference

[1] 1. ICLR-2014-OverFeat Integrated Recognition, Localization and Detection using Convolutional Networks

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment