Have you considered using transformer as a backbone？

[Wakings]

Many thanks to SuperGlobal for this work！I find that most of the current work related to instance image retrieval is based on resnet, and transformer is rarely used as the base. Is there any consideration for this? transformer is generally considered to perform better than resnet, and I also see that you use transformer instead of resnet on Google Universal Image Embedding. Thank you very much for your reply！