fundamentalvision/Deformable-DETR

Why is the classification cost in matcher formulated as the positive minus negative?

Godricly opened this issue · 1 comments

Hi, I'm confused with this part. The computation here is different from the implementation in sigmoid_focal_loss computation. What is the purpose here to use minus sign?

cost_class = pos_cost_class[:, tgt_ids] - neg_cost_class[:, tgt_ids]

found an explaination in https://arxiv.org/pdf/2304.04742.pdf.