search

Tencent / FaceDetection-DSFD

腾讯优图高精度双分支人脸检测器
Other
2.91k stars 728 forks source link

why fem is different with paper #23

Closed GGjerry01 closed 5 years ago

GGjerry01 commented 5 years ago

image image

swordlidev commented 5 years ago

Hi，this fem function is just right part of picture，which is constructed by dilation conv，the left part is fpn.

GGjerry01 commented 5 years ago

Hi，this fem function is just right part of picture，which is constructed by dilation conv，the left part is fpn.

thanks, reply. ye i am confusing with the structure of dilation conv. in paper, splitting the dimension N to N/3. each one does diffierent dilation conv and then cat it. the output is N. but in the function ,the output is 512 right?

swordlidev commented 5 years ago

Yes，the structure of dilation conv outputs feature maps，which have the same channel numbers N=512. N also can be unified to be 512 by fpn（left part）.

GGjerry01 commented 5 years ago

Yes，the structure of dilation conv outputs feature maps，which have the same channel numbers N=512. N also can be unified to be 512 by fpn（left part）.

so,it splits 512 to 256,128,128 not 512/3 right?

swordlidev commented 5 years ago

yes，N/3 just means three branches,