search

aim-uofa / AdelaiDet

AdelaiDet is an open source toolbox for multiple instance-level detection and recognition tasks.
https://git.io/AdelaiDet
Other
3.34k stars 643 forks source link

Detection branch of ABCNet #318

Open NguyenHongSon1103 opened 3 years ago

NguyenHongSon1103 commented 3 years ago

Hi ! Thank for your work ! I don't familiar with pytorch so i need to reimplement it by TF2 for my project at school. I want to ask something:

  1. Does Stack of 4 convolution layer is text detection head ? So if we use FPN, each output feature map will have a stack ?
  2. The paper said that it output 8 control points for each text which only need a 16 channel convolution layer. But what exacly is output and its shape of text detection head ? I think it can be (number of text, 16), but what happend if no text or more than 1 text in image ?

Thank for your answer !

NguyenHongSon1103 commented 3 years ago

After look at figure 2 in FCOS paper, question 1 is answer its self. But still confuse about output of detection head, maybe there are 2 feature map, (W, H, 16) for text's control points and (W, H, 2) for probability of text/background ?