Deep-Cross-Modal-Projection-Learning-for-Image-Text-Matching icon indicating copy to clipboard operation
Deep-Cross-Modal-Projection-Learning-for-Image-Text-Matching copied to clipboard

Published 20 hours ago • verified publisher icon labyrinth7x

Question about the code of CMPC loss.

Open ZhanYang-nwpu opened this issue 3 years ago • 0 comments

Hello. Thank you for your excellent work! But I noticed a problem with CMPC Loss. In the paper, "the vector projection of image feature xi onto normalized text feature ̄ zi" is not implemented according to the formula. In the code, the calculation is done directly by using the dot product. Is this code inconsistent with the writing of the paper? It is not known whether this has any impact on model performance. Could you provide some explanation? Thanks a lot!

ZhanYang-nwpu avatar Sep 29 '22 03:09 ZhanYang-nwpu