search

deepinsight / insightface

State-of-the-art 2D and 3D Face Analysis Project
https://insightface.ai
23.44k stars 5.42k forks source link

question about the retinaface #2028

Open wwdok opened 2 years ago

wwdok commented 2 years ago
  1. why the classification loss have two value, i think one value is enough, such as 0 means no face, 1 means having face.
  2. why there is a crop after unsample: image
nttstar commented 2 years ago

As there're two anchors in one feature map

wwdok commented 2 years ago

@nttstar I know there are two anchors for each grid, let me describe my question more detailedly with following image, it is one of the heads of the model: image As there are two anchors for each grid, so 20 stands for 5 keypoints with x y coordinates for 2 anchors, 8 stands for 4 box regression value for 2 anchors, then 4 stands for 2 classification value for 2 anchors ? BTW, do you konw my 2nd quesion ? Thanks in advance !