search

raoyongming / DenseCLIP

[CVPR 2022] DenseCLIP: Language-Guided Dense Prediction with Context-Aware Prompting
505 stars 38 forks source link

some question about pixel-text matching loss #53

Closed Yu-zhengbo closed 5 months ago

Yu-zhengbo commented 5 months ago

In the paper, I did't find the description about the pixel-text matching loss. Then I try to find the loss in the open source code, but I fall. In the follow picture, the self.with_identity_head and self.with_auxiliary_head are None, so I just can find the task loss. Please help me. Thank you. image

raoyongming commented 5 months ago

Hi, the pixel-text matching loss is computed using identity head since it already is a classification map and doesn't require a linear classifier head. The head is not None (see here).