Have you trained resnet50 on Google-Landmarks dataset?

[Jack-xiaoxin]

Hello. Thank you very much for releasing the scripts and sharing your great work. The performance of your work is very impressive.

I have two question:

1、we evaluate Resnet101-AP-LM18, it performed well on pitts250k and tokyo24/7 and better than models trained on Landmarks-clean. Our embedded devices have limited compute ability and space, so we prefer resnet50 as backbone, but we can't find Resnet50-AP-LM18.

Have you trained resnet50-AP on Google-Landmarks Dataset? If you do, can you release it ? This will help us a lot.

2、Besides, We found that APGeM performed not so good on indoor datasets such as InLoc. Do you have some suggestions for that? We're looking forward to your reply.

[almazan]

Hi @Jack-xiaoxin!

First of all, apologies for the late reply, and also thank you so much for your interest in our work.

1. We don't have such a model trained, and given that LM18 is already an outdated version of the dataset (Google Landmarks v2 is already available) I don't think we'll spend resources on training on it. However, I do plan to train new models on GLv2, including both ResNet-101 and ResNet-50 flavours. If you're interested on these as well I'd be happy to release them.

2. The models that we have released have been trained on outdoor data exclusively so the main problem you're facing is domain mismatch. One option would be to train a different model using indoor data: NAVER just released a new dataset for indoor localisation that could be a great fit for this: https://europe.naverlabs.com/blog/first-of-a-kind-large-scale-localization-datasets-in-crowded-indoor-spaces/. If, as you mentioned, you have limited compute resources, then I would suggest to look into domain transfer techniques to alleviate this domain mismatch problem.

[sungonce]

@almazan
Thanks for the great code and explanation!
I am eagerly looking for gldv2 trained weights and results.
What plan do you have for these results? I hope to see that things soon :D

[Chucy2020]

Hello. Thank you very much for releasing the scripts and sharing your great work. The performance of your work is very impressive.

I have two question:

1、we evaluate Resnet101-AP-LM18, it performed well on pitts250k and tokyo24/7 and better than models trained on Landmarks-clean. Our embedded devices have limited compute ability and space, so we prefer resnet50 as backbone, but we can't find Resnet50-AP-LM18.

Have you trained resnet50-AP on Google-Landmarks Dataset? If you do, can you release it ? This will help us a lot.

2、Besides, We found that APGeM performed not so good on indoor datasets such as InLoc. Do you have some suggestions for that? We're looking forward to your reply.

The download address of the pre-trained model provided by the author is invalid. Could you please give me the link to find it?,Thank you very much.