IDEA-Research / MP-Former

[CVPR 2023] Official implementation of the paper: MP-Former: Mask-Piloted Transformer for Image Segmentation
Other
115 stars 6 forks source link

origin of class embeddings #1

Closed yaohusama closed 1 year ago

yaohusama commented 1 year ago

I would like to ask, how to obtain class embeddings, but it does not seem to be mentioned in the paper.

HaoZhang534 commented 1 year ago

I would like to ask, how to obtain class embeddings, but it does not seem to be mentioned in the paper.

Class embeddings are learnable embeddings with shape (num_classes, num_dimensions) where each class has a unique embedding.

yaohusama commented 1 year ago

I know that the class embedding in the matching part is a learnable parameter, but is the GT class embedding added in the

training phase also learnable?

I just found out that it should be the same as dn-detr, which is to determine a learnable embedding matrix and look up the table from the matrix according to the category. Thank you for your guidance.

HaoZhang534 commented 1 year ago

@yaohusama Yes, your understanding is right.

skyABB commented 1 year ago

我希望这条消息能找到你,我对你为类开发的代码特别感兴趣——嵌入。我相信这段代码对我自己的项目非常有价值,并想询问您是否愿意考虑开源它。