sungonce / cvnet Goto Github PK

Hello, thank you so much for the amazing work! It inspires me a lot.

May I ask if there is a (rough) date for the code to be released? I am looking forward to this masterpiece :)

Hi, could you please release the data list of gld after your cleaning?

Can you provide training codes

Hello, could you please provide CVNet training scripts and tutorials? Thank you

Hi,
Thanks for releasing the codes for this amazing work.
I just quickly took a look at the entire codes, and it seems that the codes for GLD-v2 is not there.
Is there any plan to release the pretrained weights and codes for GLD-v2 dataset?

dear sir,
How to train Google's data set and when it can be released？

Hello sungonce, I can't download the training model, can you send to my email: [email protected]. Thank you very much!

First of all, thanks for your awesome work. The result is really impressive.

Here I felt confused about the training steps for GLDv2 clean version. In your paper, it is described as:

The global model is trained for 25 epochs (39.5M steps) for the training dataset, using a learning rate of 0.005625, and a batch size of 144.

And the GLDv2 clean set, as you wrote, has 1.58M images from 81k landmarks.

Then obviously, if you want to train 25 epochs, then that means you train (1.58/144)*25 = 0.27M steps, and for 39.5M steps, it is ~3657 epochs. So which description is correct, or did I just miss something?

I noticed that you are using the same model as DOLG. In your code, an updated resnet pre-training model is used, which performs better on ImageNet dataset (BRG pre-training model provided by facebook). And the existing methods, such as GeM, and SOLAR, use the one provieded by Filip Radenovic. I think it is an unfair comparison and you need to compare on a consistent benchmark to truly show the effectiveness of your method. The performance of your method on the R1M dataset is not convincing. I also pointed this out in the DOLG repository as well, but didn't get a reply. I think you guys need to clarify this matter.

Hi @sungonce, thanks for your impressive works and sharing!
I notice that you use the "torch.bmm( )" as the correlation process in "compute_crossscale_correlation( )" function. However in general, the correlation process between two images is represented as the equation bellow. Using the multiplication to represent the correlation between two images seems to be uncommon. So could you please explain the reason for such design? Sincerely for your reply!