search

VDIGPKU / M2Det

M2Det: A Single-Shot Object Detector based on Multi-Level Feature Pyramid Network
MIT License
1.46k stars 318 forks source link

A question about SFAM Module #42

Closed WangTianYuan closed 5 years ago

WangTianYuan commented 5 years ago

作者你好，我用你给出的代码复现了论文的结果，非常感谢 但是我注意到config文件中SFAM默认是不使用的，然后我测试了使用SFAM的情况，发现效果没有改进，有些反而变差了，请问是什么原因呢？

qijiezhao commented 5 years ago

Hi, SFAM requires more training time to improve the accuracy. I think it's somewhat redundant. If you keep training SFAM for a longer time, then visualize the SFAM feature, you can get a surprise : ). But for the mmAP accuracy, the improvement is somewhat minor.