Questions about FEDER

[waner1998]

Hi. In FEDER, there is a loss about the edge ( loss_edge = dice_loss(preds[6], edges)*0.125 + dice_loss(preds[7], edges)*0.25 + dice_loss(preds[8], edges)*0.5 ). However, not all datasets contain data about edge information (for example, 'CAMO' ), so I had to remove this loss about edge information, but the experimental results are much different from the results in your article. May I ask how you dealt with this in your experiments?
Sincerely,

[xiongxyowo]

Hi, we directly use the results reported in the FEDER paper without reproducing their code. Unfortunately, the authors of FEDER also didn't answer the question about how to obtain the edges (see here). You may try to use cv2.findContours to generate edges from the GT masks.

[waner1998]

hhh, thank you for your efficient reply!