为何只在PNet中使用1x1卷积代替全连接？而不再RNet和ONet中使用1x1卷积呢？

AITTSMD / MTCNN-Tensorflow

Reproduce MTCNN using Tensorflow

1.51k stars 713 forks source link

为何只在PNet中使用1x1卷积代替全连接？而不再RNet和ONet中使用1x1卷积呢？ #136

Closed Edwardmark closed 6 years ago

kingrain213 commented 6 years ago

只是粗略的计算个结果出来，减少参数和计算量？

Edwardmark commented 6 years ago

@kingrain213 不是吧，计算量是一样的

choice17 commented 6 years ago

@Edwardmark, PNet does not use fully connected layer, as input of PNet is a image without fixed resolution which allows dynamic input and output resolution while RNet, ONet need fixed size input (24x24x3 and 48x48x3), you may read https://kpzhang93.github.io/MTCNN_face_detection_alignment/index.html matlab codes for details.

Edwardmark commented 6 years ago

@choice17 Thanks man.