ligang-cs / PseCo

An official implementation of the PseCo (ECCV2022)
Apache License 2.0
133 stars 19 forks source link

feature-level consistency learning #14

Closed yu1206 closed 1 year ago

yu1206 commented 2 years ago

特征一致性学习 在训练的过程中怎么使用的呢?

Stephen-Hao commented 2 years ago

我理解就是把两个尺度相同的feature进行类似的L2约束吧?

ligang-cs commented 1 year ago

不好意思,文章解释得不太清楚。view 2是对view 1进行2倍下采样得到的,是低分辨率的图片。而检测器对低分辨率的图片,检测性能向来较差。通过Feature-level的对齐,即,使用相同的pseudo box来同时监督view 2的FPN features和view 1的FPN features,我们希望低分辨率图片(view 2),可以产生和高分辨率图片(view 1)一样强的FPN features. 具体的实现麻烦参考代码:PseCo_frcnn.py.